HUGIN API

REFERENCE MANUAL
Version 8.3

HUGIN API Reference Manual

Copyright 19902016 by Hugin Expert A/S.
All Rights Reserved.

This manual was prepared using the LATEX Document Preparation System
and the PDFTEX typesetting software.
Set in 11 point Bitstream Charter, Bitstream Courier, and AMS Euler.
Version 8.3, February 2016.

UNIX

is a trademark of The Open Group

POSIX is a trademark of IEEE

Sun, Solaris, Sun Enterprise, and Java are trademarks or registered trademarks of
Sun Microsystems, Inc. in the United States and in other countries.
Microsoft, Visual C++, Windows, Windows 95, and Windows NT are trademarks
or registered trademarks of Microsoft Corporation in the United States and/or
other countries.
Mac OS is a registered trademark of Apple Computer, Inc.
X Window System is a trademark of X Consortium, Inc.
Bitstream and Bitstream Charter are registered trademarks of Bitstream Inc.
Hugin Expert, the Hugin Raven Logo, Hugin API, Hugin Regular, and Hugin
Runtime are trademarks of Hugin Expert A/S.

Preface
The HUGIN API 8.3 Reference Manual provides a reference for the C language Application Program Interface to the HUGIN system. However, brief
descriptions of the Java and C++ versions are also provided (see Chapter 1).
The present manual assumes familiarity with the methodology of Bayesian
belief networks and (limited memory) influence diagrams (LIMIDs) as well
as knowledge of the C programming language and programming concepts.
As introductions to Bayesian belief networks and influence diagrams, the
books by Jensen and Nielsen [14] and Kjrulff and Madsen [18] are recommended. Deeper treatments of the subjects are given in the books by Cowell
et al [8], Darwiche [9], Koller and Friedman [19], and Pearl [31].

Overview of the manual

Chapter 1 explains how to use the HUGIN API within your own applications.
It also gives some general information on the functions and data types defined by the HUGIN API and explains the mechanisms for error handling.
Finally, instructions on how to take advantage of multi-processor systems to
speed up inference are given.
Chapter 2 describes the functions for creating and modifying belief networks
and LIMIDs, collectively known as domains. It also explains how to save and
load domains to/from knowledge base files.
Chapter 3 provides the tools for constructing object-oriented belief network
and LIMID models. Moreover, a function for converting an object-oriented
model to an equivalent domain is given (which is needed because inference
cannot be performed directly in an object-oriented model).
Chapter 4 describes how to construct dynamic models (Dynamic Bayesian
Networks). The concepts of temporal clone, time window, and prediction are
explained.
Tables are used to represent conditional probability distributions, policies,
utility functions, sets of experience counts, and sets of fading factors associated with the nodes of the network, as well as joint probability distributions
and so-called mixture marginals (representing marginal distributions of
iii

continuous nodes). Chapter 5 explains how to access and modify the contents of tables.
Chapter 6 describes how the contents of a conditional probability, a policy,
or a utility table can be generated from a mathematical description of the
relationship between a node and its parents.
Chapter 7 explains how to transform a domain into a secondary structure
(a junction forest), suitable for inference. This transformation is known
as compilation. It also explains how to improve performance of inference
by controlling the triangulation step and by performing approximation and
compression.
Chapter 8 explains how to access the collection of junction trees of a compiled domain and how to traverse a junction tree.
Chapter 9 shows how to handle the beliefs and the evidence that form the
core of the reasoning process in the HUGIN inference engine. This chapter
explains how to enter and retract evidence, how to determine independence
properties induced by evidence and network structure, how to retrieve beliefs and expected utilities, how to compute values of function nodes, how
to examine evidence, and how to save evidence as a case file for later use.
Chapter 10 documents the functions used to control the inference engine
itself. The chapter also explains how to perform conflict analysis, simulation,
value of information analysis, sensitivity analysis, and how to find the most
probable configurations of a set of nodes.
Chapter 11 explains how to adapt conditional probability distributions to
new evidence, and Chapter 12 describes how the network structure and the
conditional probability distributions can be extracted (learned) from data
(a set of cases).
Chapter 13 describes the NET language, a language used to specify the
nodes and the structure of a network as well as the numerical data required
to form a complete specification.
Chapter 14 describes the data set a tool that aids in loading of data provided as so-called CSV files (short for comma-separated-values files). After
the data has been loaded (as pure text), it can be modified (if necessary) and
used as case data for the learning algorithms.
Chapter 15 describes how to enter and modify information that is purely
descriptive. This information is not used by other parts of the HUGIN API.
It is used by the HUGIN GUI application to generate a graphical display of a
network.
Appendix A gives an example of a network using CG variables. Appendix B
provides a history of news and changes for all releases of the HUGIN API
since version 2.
Finally, an index is provided. The index contains the names of all functions,
types, and constants of enumeration types, defined in this manual.
iv

A note on the API function descriptions

The description of functions in this manual are given in terms of ISO/ANSI C
function prototypes, giving the names and types of the functions and their
arguments.
The notation h domain compile(103) is used to refer to an API function (in
this case, the h domain compile function). The parenthesized, superscripted
number (103) refers to the page where the function is described.

A note on the examples

Throughout the manual, brief examples are provided to illustrate the use
of particular functions. These examples will not be complete applications.
Rather, they will be small pieces of code that show how a function (or a
group of functions) might be used.
While each example is intended to illustrate the use of a particular function,
other functions of the HUGIN API will be used to make the examples more
realistic. As this manual is not intended as a tutorial but as a reference,
many examples will use functions described later in the manual. Therefore, if you read the manual sequentially, you cannot expect to be able to
understand all examples the first time through.
For the sake of brevity, most examples do not include error checking. It
should be pointed out that using this practice in real applications is strongly
discouraged.

Acknowledgements
Lars P. Fischer wrote the HUGIN API 1.1 manual, and Per Abrahamsen wrote
the HUGIN API 1.2 (Extensions) manual. The present document is partly
based on these manuals.
I would also like to thank Anders L. Madsen, Martin Karlsen, Uffe B. Kjrulff,
Marianne Bangs, Sren L. Dittmer, Michael Lang, Lars Nielsen, Lars Bo
Nielsen, and Kristian G. Olesen for providing constructive comments and
other contributions that have improved this manual as well as the API itself.
In particular, Anders L. Madsen wrote a large part of Chapter 11.
Any errors and omissions remaining in this manual are, however, my responsibility.
The development of the functionality concerning the (real-valued) function
node type (introduced in HUGIN 7.3) has been sponsored by Danish mortgage credit institution Nykredit Realkredit (www.nykredit.dk).
The development of the functionality concerning the discrete function node
type as well as the aggregate and probability operators (introduced in
v

HUGIN 7.7) has been sponsored by the research project Operational risk
in banking and finance. The project is dedicated to strengthening management of operational risk in the banking and finance sector, hereunder
develop Basel II compliant operational risk measurement and management
tools in accordance with the Advanced Measurement Approach (AMA). The
project is financed by the University of Stavanger and a consortium of Norwegian banks consisting of Sparebank 1 SR-Bank, Sparebank 1 SNN, Sparebank 1 SMN, Sparebanken Hedmark, and Sparebank 1 Oslo and Akershus.
Frank Jensen
Hugin Expert A/S
February, 2016

vi

Contents
Preface

iii

1 General Information

1.1

Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.2

Using the HUGIN API on UNIX platforms . . . . . . . . . . .

1.3

Using the HUGIN API on Windows platforms . . . . . . . . .

1.4

Naming conventions

1.5

Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.6

Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

. . . . . . . . . . . . . . . . . . . . . . 15

1.6.1 Handling errors . . . . . . . . . . . . . . . . . . . . . 19

1.6.2 General errors . . . . . . . . . . . . . . . . . . . . . . 20
1.7

Taking advantage of multiple processors . . . . . . . . . . . 21

1.7.1 Multiprocessing in the Solaris Operating Environment 22
1.7.2 Multiprocessing on Windows platforms . . . . . . . . 23

1.8

Using the HUGIN API in a multithreaded application . . . . . 23

2 Nodes and Domains

2.1

27

Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.1.1 Node category . . . . . . . . . . . . . . . . . . . . . . 27
2.1.2 Node kind . . . . . . . . . . . . . . . . . . . . . . . . 28

2.2

Domains: Creation and deletion . . . . . . . . . . . . . . . . 29

2.3

Nodes: Creation and deletion . . . . . . . . . . . . . . . . . 30

2.4

The links of the network . . . . . . . . . . . . . . . . . . . . 31

2.4.1 The requisite parents and ancestors of decision nodes 36

2.5

The number of states of a discrete node . . . . . . . . . . . . 38

2.6

The table of a node . . . . . . . . . . . . . . . . . . . . . . . 39

2.7

The name of a node . . . . . . . . . . . . . . . . . . . . . . . 42

2.8

Iterating through the nodes of a domain

vii

. . . . . . . . . . . 43

2.9

User data . . . . . . . . . . . .
2.9.1 Arbitrary user data . .
2.9.2 User-defined attributes
2.10 HUGIN Knowledge Base files .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

44
44
45
47

.
.
.
.
.
.
.
.
.
.
.
.
.

51
51
52
52
52
53
53
54
55
57
58
60
61
62

.
.
.
.
.

63
63
64
65
67
68

.
.
.
.
.

69
69
72
73
73
74

6 Generating Tables
6.1 Subtyping of discrete nodes . . . . . . . . . . . . . . . . . .
6.2 Expressions . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3 Syntax of expressions . . . . . . . . . . . . . . . . . . . . . .

77
77
78
85

3 Object-Oriented Belief Networks and LIMIDs

3.1 Classes and class collections . . . . . . .
3.2 Creating classes and class collections . .
3.3 Deleting classes and class collections . .
3.4 Naming classes . . . . . . . . . . . . . .
3.5 Creating basic nodes . . . . . . . . . . .
3.6 Naming nodes . . . . . . . . . . . . . . .
3.7 The interface of a class . . . . . . . . . .
3.8 Creating instances of classes . . . . . . .
3.9 Putting the pieces together . . . . . . . .
3.10 Creating a runtime domain . . . . . . . .
3.11 Node iterator . . . . . . . . . . . . . . .
3.12 User data . . . . . . . . . . . . . . . . . .
3.13 Saving class collections as HKB files . . .

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

4 Dynamic Bayesian Networks

4.1 Temporal clones . . . . . . . . . . . . . . .
4.2 DBN runtime domains . . . . . . . . . . .
4.3 Inference in DBNs . . . . . . . . . . . . . .
4.4 Prediction . . . . . . . . . . . . . . . . . .
4.5 The Boyen-Koller approximation algorithm
5 Tables
5.1 What is a table? . . . . . . . . . . . .
5.2 The nodes and the contents of a table
5.3 Deleting tables . . . . . . . . . . . . .
5.4 The size of a table . . . . . . . . . . .
5.5 Rearranging the contents of a table .

viii

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

6.4

Creating and maintaining models . . . . . . . . . . . . . . . 86

6.5

State labels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88

6.6

State values . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

6.7

Statistical distributions . . . . . . . . . . . . . . . . . . . . . 90
6.7.1 Continuous distributions . . . . . . . . . . . . . . . . 91
6.7.2 Discrete distributions . . . . . . . . . . . . . . . . . . 93

6.8

Generating tables . . . . . . . . . . . . . . . . . . . . . . . . 95

6.9

How the computations are done . . . . . . . . . . . . . . . . 97

7 Compilation

101

7.1

What is compilation? . . . . . . . . . . . . . . . . . . . . . . 101

7.2

Compilation . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7.3

Triangulation . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7.4

Getting a compilation log . . . . . . . . . . . . . . . . . . . . 109

7.5

Uncompilation . . . . . . . . . . . . . . . . . . . . . . . . . . 110

7.6

Compression . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

7.7

Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . 111

8 Cliques and Junction Trees

115

8.1

Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

8.2

Junction trees . . . . . . . . . . . . . . . . . . . . . . . . . . 116

8.3

Cliques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

8.4

Traversal of junction trees . . . . . . . . . . . . . . . . . . . 118

9 Evidence and Beliefs

9.1

119

Evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
9.1.1 Discrete evidence . . . . . . . . . . . . . . . . . . . . 119
9.1.2 Continuous evidence . . . . . . . . . . . . . . . . . . 121
9.1.3 Evidence in LIMIDs . . . . . . . . . . . . . . . . . . . 121

9.2

Entering evidence . . . . . . . . . . . . . . . . . . . . . . . . 121

9.3

Retracting evidence . . . . . . . . . . . . . . . . . . . . . . . 122

9.4

Determining independence properties . . . . . . . . . . . . . 123

9.5

Retrieving beliefs . . . . . . . . . . . . . . . . . . . . . . . . 125

9.6

Retrieving expected utilities . . . . . . . . . . . . . . . . . . 127

9.7

Computing function values . . . . . . . . . . . . . . . . . . . 128

9.8

Examining the evidence . . . . . . . . . . . . . . . . . . . . . 129

9.9

Case files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

ix

10 Inference

133

10.1 Propagation methods . . . . . . . . . . . . . . . . . . . . . . 133

10.1.1 Summation and maximization . . . . . . . . . . . . . 133
10.1.2 Evidence incorporation mode . . . . . . . . . . . . . 134
10.2 Propagation . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
10.3 Inference in LIMIDs: Computing optimal policies . . . . . . . 137
10.4 Conflict of evidence . . . . . . . . . . . . . . . . . . . . . . . 138
10.5 The normalization constant . . . . . . . . . . . . . . . . . . . 140
10.6 Initializing the inference engine . . . . . . . . . . . . . . . . 141
10.7 Querying the state of the inference engine . . . . . . . . . . 143
10.8 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
10.9 Value of information analysis . . . . . . . . . . . . . . . . . . 146
10.10 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . 148
10.11 Most probable configurations . . . . . . . . . . . . . . . . . . 154
10.12 Explanation . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
11 Sequential Updating of Conditional Probability Tables

159

11.1 Experience counts and fading factors . . . . . . . . . . . . . 159

11.2 Updating tables . . . . . . . . . . . . . . . . . . . . . . . . . 162
12 Learning Network Structure and Conditional Probability Tables 165
12.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
12.2 Data files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
12.3 Scoring of graphical models . . . . . . . . . . . . . . . . . . 171
12.4 Learning network structure . . . . . . . . . . . . . . . . . . . 173
12.5 Domain knowledge . . . . . . . . . . . . . . . . . . . . . . . 174
12.6 Learning conditional probability tables . . . . . . . . . . . . 175
13 The NET Language

181

13.1 Overview of the NET language . . . . . . . . . . . . . . . . . 182

13.2 Basic nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
13.3 Class instances . . . . . . . . . . . . . . . . . . . . . . . . . . 185
13.4 Temporal clones . . . . . . . . . . . . . . . . . . . . . . . . . 187
13.5 The structure of the model . . . . . . . . . . . . . . . . . . . 187
13.6 Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
13.6.1 Direct specification of the numbers . . . . . . . . . . 190
13.6.2 Using the table generation facility . . . . . . . . . . . 192
x

13.6.3 Parameter learning . . . . . . . . . . . . . . . . . . . 195

13.7 Global information . . . . . . . . . . . . . . . . . . . . . . . 196
13.8 Lexical matters

. . . . . . . . . . . . . . . . . . . . . . . . . 197

13.9 Parsing NET files . . . . . . . . . . . . . . . . . . . . . . . . . 198

13.10 Saving class collections, classes, and domains as NET files . . 201
14 Data Sets and CSV Files

203

14.1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

14.2 CSV files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
15 Display Information

209

15.1 The label of a node . . . . . . . . . . . . . . . . . . . . . . . 209

15.2 The position of a node . . . . . . . . . . . . . . . . . . . . . 210
15.3 The size of a node . . . . . . . . . . . . . . . . . . . . . . . . 210
A Belief networks with Conditional Gaussian variables

213

B HUGIN API Revision History

215

Bibliography

235

Index

239

xi

xii

Chapter 1

General Information
This chapter explains how to use the HUGIN API within your own applications. It also gives some general information on the functions and data types
defined by the HUGIN API and explains the mechanisms for error handling.
Finally, instructions on how to take advantage of multi-processor systems to
speed up inference is given.

1.1

Introduction

The HUGIN API contains a high performance inference engine that can be
used as the core of knowledge based systems built using Bayesian belief networks or limited memory influence diagrams (LIMIDs) [24]. A knowledge
engineer can build knowledge bases that model the application domain, using probabilistic descriptions of causal relationships in the domain. Given
this description, the HUGIN inference engine can perform fast and accurate
reasoning.
The HUGIN API is provided in the form of a library that can be linked into
applications written using the C, C++, or Java programming languages. The
C version provides a traditional function-oriented interface, while the C++
and Java versions provide an object-oriented interface. The present manual
describes the C interface. The C++ and Java interfaces are described in
online documentation supplied with the respective libraries.
On Windows platforms only, C# and Visual Basic language interfaces are
also available.
Additionally, a web service version of the HUGIN API is provided. This
allows the decision engine to be exercised from any programming language
using the well-known HTTP protocol.
The HUGIN API is used just like any other library. It does not require any
special programming techniques or program structures. The HUGIN API
1

does not control your application. Rather, your application controls the
HUGIN API by telling it which operations to perform. The HUGIN inference
engine sits passive until you engage it.
Applications built using the HUGIN API can make use of any other library
packages such as database servers, GUI toolkits, etc. The HUGIN API itself
only depends on (in addition to the Standard C library) the presence of the
Zlib library (www.zlib.net), which is preinstalled in the Solaris, Linux,
and Mac OS X operating environments.

1.2

Using the HUGIN API on UNIX platforms

The first step in using the C version of the HUGIN API is to include the
definitions for the HUGIN functions and data types in the program. This is
done by inserting the following line at the top of the program source code:
# include "hugin.h"
The hugin.h header file contains all the definitions for the API.
When compiling the program, you must inform the C compiler where the
header file is stored. Assuming the HUGIN system has been installed in the
directory /usr/local/hugin, the following command is used:
cc -I/usr/local/hugin/include -c myapp.c
This will compile the source code file myapp.c and store the result in the
object code file myapp.o, without linking. The -I option adds the directory
/usr/local/hugin/include to the search path for include files.
If you have installed the HUGIN system somewhere else, the path above
must be modified as appropriate. If the environment variable HUGINHOME
has been defined to point to the location of the HUGIN installation, the
following command can be used:
cc -I$HUGINHOME/include -c myapp.c
Using the environment variable, HUGINHOME, has the advantage that if the
HUGIN system is moved, only the environment variable must be changed.
When the source code, possibly stored in several files, has been compiled,
the object files must be linked to create an executable file. At this point, it
is necessary to specify that the object files should be linked with the HUGIN
library:
cc myapp.o other.o -L$HUGINHOME/lib -lhugin -lm -lz
The -L$HUGINHOME/lib option specifies the directory to search for the
HUGIN libraries, while the -lhugin option specifies the library to link
2

with. The -lz option directs the compiler/linker to link with the Zlib library (www.zlib.net). This option is needed if either of the h domain
save as kb(49) or h kb load domain(49) functions is used.
If the source code for your application is a single file, you can simplify the
above to:
cc -I$HUGINHOME/include myapp.c
-L$HUGINHOME/lib -lhugin -lm -lz -o myapp
compiling the source code file myapp.c and storing the final application in
the executable file myapp. (Note that the above command should be typed
as a single line.)
Following the above instructions will result in an executable using the singleprecision version of the HUGIN API library. If, instead, you want to use
the double-precision version of the HUGIN API library, you must define H
DOUBLE when you invoke the compiler, and specify -lhugin2 for the linking step:
cc -DH_DOUBLE -I$HUGINHOME/include myapp.c
-L$HUGINHOME/lib -lhugin2 -lm -lz -o myapp
(Again, all this should be typed on one line.)
The above might look daring, but it would typically be done in a Makefile
so that you will only have to do it once for each project.
The instructions above also assume that you are using the 32-bit version of
the HUGIN API. If, instead, you are using the 64-bit version, you will need
to specify $HUGINHOME/lib64 as the directory to search for libraries.
The hugin.h header file has been designed to work with both ISO C compliant compilers and C++ compilers. For C++ compilers, the hugin.h header
file depends on the symbol __cplusplus being defined (this symbol should
be automatically defined by the compiler).
Some API functions take pointers to stdio FILE objects as arguments. This
implies that inclusion of hugin.h also implies inclusion of <stdio.h>.
Moreover, in order to provide suitable type definitions, the standard C header
<stddef.h> is also included.

Object-oriented versions of the HUGIN API: Java and C++

The standard HUGIN API as defined by the hugin.h header file and described in the present manual has a function-oriented interface style. Objectoriented versions, more appropriate for use in object-oriented language environments, have been developed for the Java and C++ languages. These
versions have almost identical interfaces, and it should be very easy for developers to switch between them (if this should ever be necessary).
3

The Java and C++ versions use classes for modeling domains, nodes, etc.
Each of the classes has a set of methods enabling you to manipulate objects
of the class. These methods will throw exceptions when errors occur. The
exception classes are all subclasses of the main HUGIN exception class (ExceptionHugin). In Java, this is an extension of the standard Java Exception
class.
The classes, methods, and exceptions are all specified in the online documentation distributed together with these interfaces.
C++ To use the C++ HUGIN API definitions in your code, you must include
the hugin header file (note that there is no suffix):
# include "hugin"
All entities defined by the C++ API are defined within the HAPI namespace.
To access these entities, either use the HAPI:: prefix or place the following
declaration before the first use of C++ API entities (but after the hugin
header file has been included):
using namespace HAPI;
Like the C API, the C++ API is available in two versions: a single-precision
version and a double-precision version. To use the single-precision version,
use a command like the following for compiling and linking:
g++ -I$HUGINHOME/include myapp.c
-L$HUGINHOME/lib -lhugincpp -lm -lz -o myapp
(This should be typed on one line.) To use the double-precision version,
define the H DOUBLE preprocessor symbol and specify -lhugincpp2 for
the linking step:
g++ -DH_DOUBLE -I$HUGINHOME/include myapp.c
-L$HUGINHOME/lib -lhugincpp2 -lm -lz -o myapp
(Again, this should be typed on one line.)
Also, the C++ API is available as both a 32-bit and a 64-bit version. To use
the 64-bit version, specify $HUGINHOME/lib64 as the directory to search
for libraries.
Java The Java version of the HUGIN API library is provided as two files
(the files comprising the 64-bit version have an additional -64 suffix):
hapi83.jar (hapi83-64.jar) contains the Java interface to the
underlying C library. This file must be mentioned by the CLASSPATH
environment variable.
4

 libhapi83.so (libhapi83-64.so) contains the native version of

the HUGIN API for the platform used. When running the Java VM, this
file must be located in a directory mentioned by the LD_LIBRARY_PATH
environment variable.
The Java version of the HUGIN API is a double-precision library.
Java SE, version 1.4.2 or newer, is required: http://www.oracle.com/
technetwork/java/javase/downloads/.

Web Service version of the HUGIN API

The HUGIN Web Service API is the HUGIN decision engine turned into a
web service, thus allowing the decision engine to be exercised from any
programming language using the well-known HTTP protocol.
In this version, the HUGIN decision engine is implemented as a web server.
The web server responds to HTTP requests, exercising the decision engine
accordingly. The server is implemented in Java and depends on the HUGIN
Java API.
JavaScript libraries are also included for scripting the decision engine from
a web browser.
Please refer to the separate online documentation for information on how to
use the HUGIN Web Service API and to obtain coding examples for various
languages.

1.3

Using the HUGIN API on Windows platforms

C, C++, Java, C# (.NET), and Visual Basic language interfaces for the HUGIN
API are available on the Windows platforms.
Five1 sets of library files are provided with developer versions of HUGIN
for Windows: One for Microsoft Visual Studio 6.0, one for Microsoft Visual
Studio .NET 2003, one for Microsoft Visual Studio 2005, one for Microsoft
Visual Studio 2008, and one for Microsoft Visual Studio 2010.2 Each set of
library files contains libraries for all combinations of the following:
C and C++ programming languages;
Debug and Release configurations;
single-precision and double-precision.
1

For 64-bit packages, only three sets of library files are provided (for Microsoft Visual
Studio 2005, Microsoft Visual Studio 2008, and Microsoft Visual Studio 2010).
2
If you need libraries for other development environments, please contact info@hugin.
com.

Each of these libraries has two parts: An import library and a DLL. For example, the import library for the 32-bit, double-precision, C++ version of
the HUGIN API compiled for Microsoft Visual Studio 2008, Debug configuration, is named hugincpp2-8.3-vc9d.lib, and the corresponding DLL
file is named hugincpp2-8.3-vc9d.dll. For 64-bit versions, there is an
additional suffix: hugincpp2-8.3-vc9d-x64 plus the usual .lib and
.dll extensions.
In general, the library files have unique names, indicating language (C or
C++), API version number, compiler, configuration, and 32-bit/64-bit. This
naming scheme makes it possible for all DLLs to be in the search path simultaneously.

C version of the HUGIN API

The C version of the HUGIN API is located in the HDE8.3C subdirectory of
the main HUGIN installation directory.
To use the C version of the HUGIN API in your code, you must include the
hugin.h header file:
# include "hugin.h"
Microsoft Visual Studio 6.0
Let hPathi denote the main directory of the Hugin installation (for example:
C:\Program Files\Hugin Expert\Hugin Developer 8.3), and let
hConfigurationi denote the active configuration (either Debug or Release).
(1) Click Settings on the Project menu. Click the C/C++ tab, and
select Preprocessor in the Category box.
(a) Add hPathi\HDE8.3C\Include to the Additional include directories box.
(2) Click Settings on the Project menu. Click the Link tab, and select
Input in the Category box.
(a) Add hPathi\HDE8.3C\Lib\VC6\hConfigurationi to the Additional library path box.
(b) Add the import library to the Object/library modules list:
If hConfigurationi is Debug, add hugin-8.3-vc6d.lib.
If hConfigurationi is Release, add hugin-8.3-vc6.lib.
(3) Click Settings on the Project menu. Click the C/C++ tab, and
select Code Generation in the Category box.
6

 If hConfigurationi is Debug, make sure Debug Multithreaded

DLL is selected in the Use run-time library box.
If hConfigurationi is Release, make sure Multithreaded DLL is
selected in the Use run-time library box.
The above steps set up Microsoft Visual Studio 6.0 to use the single-precision
version of the HUGIN API. If you want to use the double-precision version,
modify the instructions as follows:
(1) Click Settings on the Project menu. Click the C/C++ tab, and
select Preprocessor in the Category box.
(b) Add H_DOUBLE to the Preprocessor definitions box.
(2) Click Settings on the Project menu. Click the Link tab, and select
Input in the Category box.
(b) Add the import library to the Object/library modules list:
If hConfigurationi is Debug, add hugin2-8.3-vc6d.lib.
If hConfigurationi is Release, add hugin2-8.3-vc6.lib.
When running the compiled program, the DLL corresponding to the import
library used in the compilation must be located in a directory mentioned in
the search path.
Microsoft Visual Studio .NET 2003
Let hPathi denote the main directory of the Hugin installation (for example:
C:\Program Files\Hugin Expert\Hugin Developer 8.3), and let
hConfigurationi denote the active configuration (either Debug or Release).
(1a) Click Properties on the Project menu. Click the C/C++ folder,
and select the General property page.
Add hPathi\HDE8.3C\Include to the Additional Include Directories property.
(2a) Click Properties on the Project menu. Click the Linker folder, and
select the General property page.
Add hPathi\HDE8.3C\Lib\VC7\hConfigurationi to the Additional Library Directories property.
(2b) Click Properties on the Project menu. Click the Linker folder, and
select the Input property page.
7

 Add the import library to the Additional Dependencies property:

If hConfigurationi is Debug, add hugin-8.3-vc7d.lib.
If hConfigurationi is Release, add hugin-8.3-vc7.lib.
(3) Click Properties on the Project menu. Click the C/C++ folder,
and select the Code Generation property page.
If hConfigurationi is Debug, make sure Multi-threaded Debug
DLL (/MDd) is selected in the Runtime Library property.
If hConfigurationi is Release, make sure Multi-threaded DLL
(/MD) is selected in the Runtime Library property.
The above steps set up Microsoft Visual Studio .NET 2003 to use the singleprecision version of the HUGIN API. If you want to use the double-precision
version, modify the instructions as follows:
(1b) Click Properties on the Project menu. Click the C/C++ folder,
and select the Preprocessor property page.
Add H_DOUBLE to the Preprocessor Definitions property.
(2b) Click Properties on the Project menu. Click the Linker folder, and
select the Input property page.
Add the import library to the Additional Dependencies property:
If hConfigurationi is Debug, add hugin2-8.3-vc7d.lib.
If hConfigurationi is Release, add hugin2-8.3-vc7.lib.
When running the compiled program, the DLL corresponding to the import
library used in the compilation must be located in a directory mentioned in
the search path.
Microsoft Visual Studio 2005, 2008, and 2010
The instructions for using the HUGIN C API in Visual Studio 2005, Visual
Studio 2008, and Visual Studio 2010 are the same as those given above for
Visual Studio .NET 2003, except that the compilers are named vc8, vc9,
and vc10 instead of vc7.
64-bit versions of the HUGIN C API are available for Visual Studio 2005,
Visual Studio 2008, and Visual Studio 2010. The library files for these versions have an additional suffix (-x64). This suffix must be included when
the import library is specified.
8

C++ object-oriented version of the HUGIN API

The C++ version of the HUGIN API is located in the HDE8.3CPP subdirectory of the main HUGIN installation directory. The documentation for
all classes and their members is located in the Doc subdirectory below the
HDE8.3CPP directory.
To use the C++ HUGIN API definitions in your code, you must include the
hugin header file (note that there is no suffix):
# include "hugin"
All entities defined by the C++ API are defined within the HAPI namespace.
To access these entities, either use the HAPI:: prefix or place the following
declaration before the first use of C++ API entities (but after the hugin
header file has been included):
using namespace HAPI;
Microsoft Visual Studio 6.0
Let hPathi denote the main directory of the Hugin installation (for example:
C:\Program Files\Hugin Expert\Hugin Developer 8.3), and let
hConfigurationi denote the active configuration (either Debug or Release).
(1) Click Settings on the Project menu. Click the C/C++ tab, and
select Preprocessor in the Category box.
(a) Add hPathi\HDE8.3CPP\Include to the Additional include directories box.
(2) Click Settings on the Project menu. Click the Link tab, and select
Input in the Category box.
(a) Add hPathi\HDE8.3CPP\Lib\VC6\hConfigurationi to the Additional library path box.
(b) Add the import library to the Object/library modules list:
If hConfigurationi is Debug, add hugincpp-8.3-vc6d.lib.
If hConfigurationi is Release, add hugincpp-8.3-vc6.lib.
(3) Click Settings on the Project menu. Click the C/C++ tab, and
select Code Generation in the Category box.
If hConfigurationi is Debug, make sure Debug Multithreaded
DLL is selected in the Use run-time library box.
If hConfigurationi is Release, make sure Multithreaded DLL is
selected in the Use run-time library box.
9

(4) Click Settings on the Project menu. Click the C/C++ tab, and
select C++ Language in the Category box.
Make sure Enable exception handling is selected.
Make sure Enable Run-Time Type Information (RTTI) is selected.
The above steps set up Microsoft Visual Studio 6.0 to use the single-precision
version of the HUGIN API. If you want to use the double-precision version,
modify the instructions as follows:
(1) Click Settings on the Project menu. Click the C/C++ tab, and
select Preprocessor in the Category box.
(b) Add H_DOUBLE to the Preprocessor definitions box.
(2) Click Settings on the Project menu. Click the Link tab, and select
Input in the Category box.
(b) Add the import library to the Object/library modules list:
If hConfigurationi is Debug, add hugincpp2-8.3-vc6d.lib.
If hConfigurationi is Release, add hugincpp2-8.3-vc6.lib.
When running the compiled program, the DLL corresponding to the import
library used in the compilation must be located in a directory mentioned in
the search path.
Microsoft Visual Studio .NET 2003
Let hPathi denote the main directory of the Hugin installation (for example:
C:\Program Files\Hugin Expert\Hugin Developer 8.3), and let
hConfigurationi denote the active configuration (either Debug or Release).
(1a) Click Properties on the Project menu. Click the C/C++ folder,
and select the General property page.
Add hPathi\HDE8.3CPP\Include to the Additional Include Directories property.
(2a) Click Properties on the Project menu. Click the Linker folder, and
select the General property page.
Add hPathi\HDE8.3CPP\Lib\VC7\hConfigurationi to the Additional Library Directories property.
(2b) Click Properties on the Project menu. Click the Linker folder, and
select the Input property page.
10

 Add the import library to the Additional Dependencies property:

If hConfigurationi is Debug, add hugincpp-8.3-vc7d.lib.
If hConfigurationi is Release, add hugincpp-8.3-vc7.lib.
(3) Click Properties on the Project menu. Click the C/C++ folder,
and select the Code Generation property page.
If hConfigurationi is Debug, make sure Multi-threaded Debug
DLL (/MDd) is selected in the Runtime Library property.
If hConfigurationi is Release, make sure Multi-threaded DLL
(/MD) is selected in the Runtime Library property.
(4a) Click Properties on the Project menu. Click the C/C++ folder,
and select the Code Generation property page
Make sure Enable C++ Exceptions is selected.
(4b) Click Properties on the Project menu. Click the C/C++ folder,
and select the Language property page
Make sure Enable Run-Time Type Info is selected.
The above steps set up Microsoft Visual Studio .NET 2003 to use the singleprecision version of the HUGIN API. If you want to use the double-precision
version, modify the instructions as follows:
(1b) Click Properties on the Project menu. Click the C/C++ folder,
and select the Preprocessor property page.
Add H_DOUBLE to the Preprocessor Definitions property.
(2b) Click Properties on the Project menu. Click the Linker folder, and
select the Input property page.
Add the import library to the Additional Dependencies property:
If hConfigurationi is Debug, add hugincpp2-8.3-vc7d.lib.
If hConfigurationi is Release, add hugincpp2-8.3-vc7.lib.
When running the compiled program, the DLL corresponding to the import
library used in the compilation must be located in a directory mentioned in
the search path.
11

Microsoft Visual Studio 2005, 2008, and 2010

The instructions for using the HUGIN C++ API in Visual Studio 2005, Visual
Studio 2008, and Visual Studio 2010 are the same as those given above for
Visual Studio .NET 2003, except that the compilers are named vc8, vc9,
and vc10 instead of vc7.
64-bit versions of the HUGIN C++ API are available for Visual Studio 2005,
Visual Studio 2008, and Visual Studio 2010. The library files for these versions have an additional suffix (-x64). This suffix must be included when
the import library is specified.

.NET version of the HUGIN API

The .NET version of the HUGIN API is located in the HDE8.3CS subdirectory
of the main HUGIN installation directory (called hPathi below).
The documentation for all classes and their members is located in the hPathi\
HDE8.3CS\Doc directory. The documentation is written for C#, but the API
can also be used with other .NET-based languages.
All entities defined by the HUGIN .NET API are defined within the HAPI
namespace. To access these entities, either use the HAPI. prefix, or place
the following declaration before any namespace and class declarations:
using HAPI;
The HUGIN .NET API is provided in the form of a DLL targeting the Microsoft .NET framework. There are eight versions of the HUGIN .NET API,
corresponding to all combinations of:
single-precision and double-precision;
32-bit and 64-bit Windows platforms;
version 2.0 and version 4.0 of the Microsoft .NET framework.
The 32-bit DLLs targeting version 2.0 of the Microsoft .NET framework are
named hugincs-8.3-2.0.dll (single-precision) and hugincs2-8.32.0.dll (double-precision), where 8.3 denotes the version number of the
HUGIN API, and 2.0 denotes the version of the .NET framework targeted by
the DLL. The names of the DLLs targeting version 4.0 of the .NET framework
have a similar structure. The 64-bit DLLs have an additional suffix (-x64).
Note that certain types vary between the DLLs: The h number t type is either a single-precision or a double-precision floating-point type, and size t
and h index t are either 32-bit or 64-bit integer types. In order to make
switching between the different HUGIN .NET API versions easier, it is recommended to declare aliases for h number t, size t and h index t in the
appropriate source files and use these aliases as types. This can be done by
placing the following piece of code at the beginning of the source files:
12

#if X64
using size_t = System.UInt64;
using h_index_t = System.Int64;
#else
using size_t = System.UInt32;
using h_index_t = System.Int32;
#endif
#if H_DOUBLE
using h_number_t = System.Double;
#else
using h_number_t = System.Single;
#endif
The symbol H_DOUBLE must be defined (only) when using a double-precision version of the HUGIN .NET API, and the symbol X64 must be defined
(only) when using a 64-bit version of the HUGIN .NET API. (See more
examples in the documentation accompanying the HUGIN .NET API.)
The Microsoft .NET Framework (version 2.0 or version 4.0) is required by
the HUGIN .NET API: http://msdn.microsoft.com/netframework/
Microsoft Visual Studio 2005, 2008, and 2010
The following steps set up a Microsoft Visual Studio (2005, 2008, and 2010)
C# Project to use the .NET version of the HUGIN API:
(1) Click Add Reference on the Project menu. Click the Browse tab,
browse to the location of the HUGIN .NET API DLL files and select a
file corresponding to the desired version.
(2) Click Properties on the Project menu. Click the Build fan, and
configure Platform target to either x86 or x64 (according to the
HUGIN .NET API version used).
(3) If using the type aliasing scheme described above, define the appropriate symbols in the field Conditional compilation symbols under the
General category on the Build fan.
When running the compiled program, the DLL referenced in the project
must be located in a directory mentioned in the search path.

COM interface version of the HUGIN API

The COM interface version of the HUGIN API (for use with, e.g., Microsoft
Excel) is based on the HUGIN .NET API. It is located in the HDE8.3COM subdirectory of the main HUGIN installation directory (called hPathi below). It
is identical in terms of classes and function names with a few exceptions:
13

 COM only has parameterless constructors. Thus, HUGIN objects must

be constructed using an instance of the special HVBA factory class.
COM does not support overloading of methods. Thus, such methods are renamed into similar but distinct names. Examples and documentation of differences between the COM interface and the normal
HUGIN .NET API is located in the hPathi\HDE8.3COM\Doc directory.
For documentation of classes and their members, refer to the HUGIN
.NET API documentation located in hPathi\HDE8.3CS\Doc. Naming
of library files follows the same conventions as the naming of the
HUGIN .NET API library files. To use the HUGIN COM interface in
a Microsoft Excel macro, perform the following step:
Click References on the Project menu, and select the desired
HUGIN library, e.g., huginCOM-8.3-2.0-x64.dll in the list
of available modules.
The HUGIN COM API replaces the HUGIN API ActiveX Server, which is now
deprecated.

Java version of the HUGIN API

The Java version of the HUGIN API is located in the HDE8.3J subdirectory
of the main HUGIN installation directory (called hPathi below).
The documentation for all the classes and their members is located in the
hPathi\HDE8.3J\Doc directory. An entry to this documentation is installed
in the Start-up menu.
When running a Hugin-based Java application, the Java VM must have access to the following files (the files comprising the 64-bit version have an
additional -64 suffix):
hapi83.jar (hapi83-64.jar): This file is located in the hPathi\
HDE8.3J\Lib directory. Add this JAR file to the classpath when running the Java VM: For the Sun Java SE VM, set the CLASSPATH environment variable to include hPathi\HDE8.3J\Lib\hapi83.jar, or
specify it using the -cp (or the -classpath) option of the java.exe
command.
hapi83.dll (hapi83-64.dll): This file is located in the hPathi\
HDE8.3J\Bin directory. When running the Java VM, this file must be
in the search path (or specified using the -Djava.library.path
option).
The Java version of the HUGIN API is a double-precision library.
Java SE, version 1.4.2 or newer, is required: http://www.oracle.com/
technetwork/java/javase/downloads/.
14

Web Service version of the HUGIN API

The HUGIN Web Service API is the HUGIN decision engine turned into a
web service, thus allowing the decision engine to be exercised from any
programming language using the well-known HTTP protocol.
In this version, the HUGIN decision engine is implemented as a web server.
The web server responds to HTTP requests, exercising the decision engine
accordingly. The server is implemented in Java and depends on the HUGIN
Java API.
JavaScript libraries are also included for scripting the decision engine from
a web browser.
The web service version of the HUGIN API is located in the HDE8.3WS subdirectory of the main HUGIN installation directory (called hPathi below).
The documentation for all features of the web service is located in the
hPathi\HDE8.3WS\Doc directory. An entry to this documentation is installed in the Start-up menu.
Please refer to the separate online documentation for information on how to
use the HUGIN Web Service API and to obtain coding examples for various
languages.

1.4

Naming conventions

Naming conventions for the C version

The HUGIN C API reserves identifiers beginning with h . Your application
should not use any such names as they might interfere with the HUGIN
API. (The HUGIN API also uses names beginning with h internally; you
shouldnt use any such names either.)
The HUGIN API uses various types for representing domains, nodes, tables,
cliques, junction trees, error codes, triangulation methods, etc.
All types defined by the HUGIN API have the suffix t.
The set of types, defined by the HUGIN API, can be partitioned into two
groups: scalar types and opaque references.

Naming conventions for the Java and C++ versions

The Java and C++ HUGIN API classes have been constructed based on the
different HUGIN API opaque pointer types (see Section 1.5). For example,
the h domain t type in C corresponds to the Domain class in Java/C++.
The convention is that you uppercase all letters following an underscore
character ( ), remove the h prefix and t (or T after uppercasing) suffix,
15

and remove all remaining underscore characters. So, for example, the following classes are defined in the Java and C++ APIs: Clique, Expression,
JunctionTree, Model, Node, and Table.
There are some differences between C and object-oriented languages such
as Java and C++ that made it natural to add some extra classes. These
include different Node subclasses (DiscreteChanceNode, DiscreteDecisionNode, BooleanDCNode, LabelledDDNode, etc.) and a lot of Expression subclasses (AddExpression, ConstantExpression, BinomialDistribution, BetaDistribution, etc.). Each group forms their own class hierarchy below the corresponding superclass. Some of the most specialized Node classes use abbreviations in their names (to avoid too long class names): e.g., BooleanDCNode
is a subclass of DiscreteChanceNode which again is a subclass of Node. Here,
BooleanDCNode is abbreviated from BooleanDiscreteChanceNode.
The methods defined on the Java/C++ HUGIN API classes all correspond
to similar C API functions. For example, the setName method of the Node
class corresponds to h node set name(42) . The rule is: the h prefix is removed, letters immediately following all (other) underscore characters are
uppercased, and, finally, the underscore characters themselves are removed.
There are some exceptions where functions correspond to class constructors:
e.g., the h domain new node(30) function in the C version corresponds to a
number of different Node subclass constructors in the Java/C++ versions.

1.5

Types

Opaque pointer types

All (structured) objects within the HUGIN API are represented as opaque
pointers. An opaque pointer is a well-defined, typed, pointer that points
to some data that is not further defined. Using opaque pointers makes it
possible to manipulate references to data, without knowing the structure of
the data itself.
This means that the HUGIN API provides pointers to these types but does
not define the structure of the data pointed at. The real data are stored in
structures, but the definitions of these structures are hidden. The reason
for this is that manipulation of these structures requires knowledge of the
workings of the inference engine, and that hiding the structure makes applications independent of the actual details, preventing that future changes
to the internals of the HUGIN API require changes in user programs.
Values of opaque pointer types should only be used in the following ways:
As sources and destinations of assignments.
As arguments to and return values from functions (both HUGIN API
and user-defined functions).
16

 In comparisons with NULL or 0 or another opaque pointer value of the

same type.
You should never try to dereference these pointers. Objects, referenced by
an opaque pointer, should only be manipulated using the API functions. This
ensures that the internal data structures are always kept consistent.

Scalar types
Probabilistic reasoning is about numbers, so the HUGIN API will of course
need to handle numbers. The beliefs and utilities used in the inference engine are of type h number t, which is defined as a single-precision floatingpoint value in the standard version of the HUGIN library. The HUGIN API
also defines another floating-point type, h double t, which is defined as a
double-precision floating-point type in the standard version of the HUGIN
API. This type is used to represent quantities that are particularly sensitive to
range (e.g., the joint probability of evidence see h domain get normalization constant(140) ) and precision (e.g., the summation operations performed
as part of a marginalization operation is done with double precision).
The reason for introducing the h number t and h double t types is to make
it easier to use higher precision versions of the HUGIN API with just a simple
recompilation of the application program with some extra flags defined.
The HUGIN API uses a number of enumeration types. Some examples: The
type h triangulation method t defines the possible triangulation methods
used during compilation; the type h error t defines the various error codes
returned when errors occur during execution of API functions. Both of these
types will have new values added as extra features are added to the HUGIN
API in the future.
Many functions return integer values. However, these integer values have
different meanings for different functions.
Functions with no natural return value simply return a status result that
indicates if the function failed or succeeded. If the value is zero, the function
succeeded; if the value is nonzero, the function failed and the value will
be the error code of the error. Such functions can be easily recognized by
having the return type h status t.
Some functions have the return type h boolean t. Such functions have truth
values (i.e., true and false) as their natural return values. These functions
will return a positive integer for true, zero for false, and a negative integer
if an error occurs. The nature of the error can be revealed by using the h
error code(18) function and friends.
The HUGIN API also defines a number of other types for general use: The
type h string t is used for character strings (this type is used for node
names, file names, labels, etc.). The type h count t is an integral type to
17

denote counts (e.g., the number of states of a node), and h index t is an

integral type to denote indexes into ordered lists (e.g., an identification of
a particular state of a node); all (non-error) values of these types are nonnegative, and a negative value from a function returning a value of one of
these types indicates an error.

1.6

Errors

Several types of errors can occur when using a function from the HUGIN
API. These errors can be the result of errors in the application program, of
running out of memory, of corrupted data files, etc.
As a general principle, the HUGIN API will try to recover from any error as
well as possible. The API will inform the application program of the problem
and take no further action. It is then up to the application program to choose
an appropriate action.
This way of error handling is chosen to give the application programmer the
highest possible degree of freedom in dealing with errors. The HUGIN API
will never make a choice of error handling, leaving it up to the application
programmer to create as elaborate an error recovery scheme as needed.
When a HUGIN API function fails, the data structures will always be left in a
consistent state. Moreover, unless otherwise stated explicitly for a particular
function, this state can be assumed identical to the state before the failed
API call.
To communicate errors to the user of the HUGIN API, the API defines the
enumeration type h error t. This type contains constants to identify the
various types of errors. All constants for values of the h error t type have
the prefix h error .
All functions in the HUGIN API (except those described in this section) set
an error indicator. This error indicator can be inspected using the h error
code function.
x

h error t h error code (void)

Return the error indicator for the most recent call to an API function (other
than h error code, h error name, and h error description). If this call was
successful, h error code will return h error none (which is equal to zero). If
this call failed, h error code will return a nonzero value indicating the nature
of the error. If no relevant API call has been made, h error code will return
h error none (but see also Section 1.8 for information on the error indicator
in multithreaded applications).
All API functions return a value. Instead of explicitly calling h error code to
check for errors, this return value can usually be used to check the status
(success or failure) of an API call.
18

All functions with no natural return value (i.e., the return type is void) have
been modified to return a value. These functions have return type h status
t (which is an alias for an integral type). A zero result from such a function
indicates success while a nonzero result indicates failure. Other functions
use an otherwise impossible value to indicate errors. For example, consider
the h node get belief (125) function which returns the belief for a state of a
(chance) variable. This is a nonnegative number (and less than or equal
to one since it is a probability). This function returns a negative number
if an error occurred. Such a convention is not possible for the h node get
expected utility(127) function since any real number is a valid utility; in this
case, the h error code function must be used.
Also, most functions that return a pointer value use NULL to indicate errors.
The only exception is the group of functions that handle arbitrary user
data (see Section 2.9.1) since NULL can be a valid datum.
It is important that the application always checks for errors. Even the most
innocent-looking function might generate an error.
Note that, if an API function returns a value indicating that an error occurred, the inference engine may be in a state where normal progress of
the application is impossible. This is the case if, say, a domain could not be
loaded. For the sanity of the application it is therefore good programming
practice to always examine return values and check for errors, just like when
using ordinary Standard C library calls.

1.6.1

Handling errors

The simplest way to deal with errors in an application is to print an error

message and abort execution of the program. To generate an appropriate
error message, the following functions can be used.
Each error has a short unique name, which can be used for short error messages.
x

h string t h error name (h error t code)

Return the name of the error with code code. If code is not a valid error code,
"no_such_error" is returned.

h string t h error description (h error t code)

Return a long description of the error with code code. This description is
suitable for display in, e.g., a message window. The string contains no newline characters, so you have to format it yourself.
Example 1.1 The following code fragment attempts to load a domain from the
HUGIN Knowledge Base file named file name. The file is assumed to be protected
by password.

19

h_domain_t d;
...
if ((d = h_kb_load_domain (file_name, password)) == NULL)
{
fprintf (stderr, "h_kb_load_domain failed: %s\n",
h_error_description (h_error_code ()));
exit (EXIT_FAILURE);
}
If the domain could not be loaded, an error message is printed and the program terminates. Lots of things could cause the load operation to fail: the file is non-existing
or unreadable, the HUGIN KB file was generated by an incompatible version of the
API, the HUGIN KB file was corrupted, insufficient memory, etc.

More sophisticated error handling is also possible by reacting to a specific

error code.
Example 1.2 The propagation functions (see Section 10.2) may detect errors that
will often not be considered fatal. Thus, more sophisticated error handling than
simple program termination is required.
h_domain_t d;
...
if (h_domain_propagate
(d, h_equilibrium_sum, h_mode_normal) != 0)
switch (h_error_code ())
{
case h_error_inconsistency_or_underflow:
/* impossible evidence has been detected,
retract some evidence and try again */
...
break;
...
default:
...
}

1.6.2

General errors

Here is a list of some error codes that most functions might generate.
h error usage This error code is returned when a trivial violation of the
interface for an API function has been detected. Examples of this error:
NULL pointers are usually not allowed as arguments (if they are, it will
be stated so explicitly); asking for the belief in a non-existing state of
a node; etc.
20

h error no memory The API function failed because there was insufficient
(virtual) memory available to perform the operation.
h error io Functions that involve I/O (i.e., reading from and writing to files
on disk). The errors could be: problems with permissions, files do not
exist, disk is full, etc.

1.7

Taking advantage of multiple processors

In order to achieve faster inference through parallel execution on multiprocessor systems, many of the most time-consuming table operations have
been made threaded. Note, however, that in the current implementation
table operations for compressed domains (see Section 7.6) are not threaded.
The creation of threads (or tasks) is controlled by two parameters: the desired level of concurrency and the grain size. The first of these parameters
specifies the maximum number of threads to create when performing a specific table operation, and the second parameter specifies a lower limit on
the size of the tasks to be performed by the threads. The size of a task is
approximately equal to the number of floaing-point operations needed to
perform the task (e.g., the number of elements to sum when performing a
marginalization task).
h status t h domain set concurrency level
(h domain t domain, size t level)

This function sets the level of concurrency associated with domain to level
(this must be a positive number). Setting the concurrency level parameter to 1 will cause all table operations (involving tables originating from
domain) to be performed sequentially. The initial value of this parameter
is 1.
Note that the concurrency level and the grain size parameters are specific to
each domain.3 Hence, the parameters must be explicitly set for all domains
for which parallel execution is desired.
h count t h domain get concurrency level (h domain t domain)

This function returns the current concurrency level associated with domain.
h status t h domain set grain size (h domain t domain, size t size)

This function sets the grain size parameter associated with domain to size
(this value must be positive). The initial value of this parameter is 10 000.
3

Chapter 2 explains the domain concept as used in the HUGIN API.

21

h count t h domain get grain size (h domain t domain)

This function returns the current value of the grain size parameter associated with domain.
A table operation involving discrete nodes can naturally be divided into a
number (n) of suboperations corresponding to the state values of one or
more of the discrete nodes. These suboperations are distributed among a
number (m) of threads such that each thread performs either bn/mc or
dn/me suboperations. The number m of threads is chosen to be the highest
number satisfying m l and m n dg/se, where l is the concurrency
level, s is the suboperation size, and g is the grain size. If no number m 2
satisfies these conditions, the table operation is performed sequentially.

1.7.1

Multiprocessing in the Solaris Operating Environment

In order to take advantage of multi-processor systems running the Solaris

Operating Environment, you must link your application with a threads library:
cc myapp.o otherfile.o
-L$HUGINHOME/lib -lhugin -lm -lz -lpthread
(This should be typed on one line.)
If you omit the -lpthread linker option, you get a single-threaded executable.
Note that the Solaris Operating Environment provides two threads libraries:
libpthread for POSIX threads and libthread for Solaris threads. The
(Solaris version of the) HUGIN API can be used with both of these libraries.4
Due to the nature of the Solaris scheduling system and the fact that the
threads created by the HUGIN API are compute-bound, it is necessary (that
is, in order to take advantage of multiple processors) to declare how many
threads should be running at the same time.
This is done using the POSIX threads function pthread setconcurrency (or the
Solaris threads function thr setconcurrency).
Example 1.3 Assuming we have a system (running the Solaris Operating Environment) with four processors (and we want to use them all for running our application), we tell the HUGIN API to create up to four threads at a time, and we tell the
Solaris scheduler to run four threads simultaneously.
# include "hugin.h"
# include <pthread.h>
...
4
Experiments performed on a Sun Enterprise 250 running Solaris 7 (5/99) indicates that
the best performance is achieved using the POSIX threads library.

22

h_domain_t d;
...
h_domain_set_concurrency_level (d, 4);
pthread_setconcurrency (4);
...
/* do compilations, propagations, and other stuff
that involves inference */
We could use thr setconcurrency instead of pthread setconcurrency (in that case we
would include <thread.h> instead of <pthread.h>).

1.7.2

Multiprocessing on Windows platforms

The HUGIN API can only be used in multithreaded mode on Windows

platforms, so nothing special needs to be done for multiprocessing. (See
Section 1.3 for instructions on how to use the HUGIN API in a Windows
application.)

1.8

Using the HUGIN API in a multithreaded

application

The HUGIN API can be used safely in a multithreaded application. The major obstacle to thread-safety is shared data for example, global variables.
The only global variable in the HUGIN API is the error code variable. When
the HUGIN API is used in a multithreaded application, an error code variable is maintained for each thread. This variable is allocated the first time
it is accessed. It is recommended that the first HUGIN API function (if any)
being called in a specific thread be the h error code(18) function. If this function returns zero, it is safe to proceed (i.e., the error code variable has been
successfully allocated). If h error code returns nonzero, the thread must not
call any other HUGIN API function, since the HUGIN API functions critically
depend on being able to read and write the error code variable. (Failure to
allocate the error code variable is very unlikely, though.)
Example 1.4 This code shows the creation of a thread, where the function executed by the thread calls h error code(18) as the first HUGIN API function. If this
call returns zero, it is safe to proceed.
This example uses POSIX threads.
# include "hugin.h"
# include <pthread.h>
pthread_t thread;
void *data; /* pointer to data used by the thread */
void *thread_function (void *data)

23

{
if (h_error_code () != 0)
return NULL; /* it is not safe to proceed */
/* now the Hugin API is ready for use */
...
}
...
pthread_create (&thread, NULL, thread_function, data);
Note that the check for h error code(18) returning zero should also be performed
for the main (only) thread in a multithreaded (singlethreaded) application, when
using a thread-safe version of the HUGIN API (all APIs provided by Hugin Expert
A/S is thread-safe as of version 6.1).

In order to create a multithreaded application, it is necessary to link with a

thread library. See the previous section for instructions on how to do this.
(You most likely also need to define additional compiler flags in order to get
thread-safe versions of functions provided by the operating system see the
system documentation for further details.)
The most common usage of the HUGIN API in a multithreaded application will most likely be to have one or more dedicated threads to process
their own domains (e.g., insert and propagate evidence, and retrieve new
beliefs). In this scenario, there is no need (and is also unnecessarily inefficient) to protect each node or domain by a mutex (mutual exclusion)
variable, since only one thread has access to the domain. However, if there
is a need for two threads to access a common domain, a mutex must be
explicitly used.
Example 1.5 The following code fragment shows how a mutex variable is used
to protect a domain from being accessed by more than one thread simultaneously.
(This example uses POSIX threads.)
# include "hugin.h"
# include <pthread.h>
h_domain_t d;
pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
...
/* In Thread A: */
if (pthread_mutex_lock (&mutex) != 0)
/* handle error */ ...;
else
{
/* now domain d can be used;
for example, evidence can be entered and
propagated, and beliefs can be retrieved;
or, the network can be modified; etc. */

24

...
pthread_mutex_unlock (&mutex);
}
...
/* In Thread B: */
if (pthread_mutex_lock (&mutex) != 0)
/* handle error */ ...;
else
{
/* use domain d */
...
pthread_mutex_unlock (&mutex);
}
Since domain d is being used by more than one thread, it is important that while
one thread is modifying the data structures belonging to d, other threads do not
attempt to read or write the same data structures. This is achieved by requiring all
threads to lock the mutex variable while they access the data structures of d. The
thread library ensures that only one thread at a time can lock the mutex variable.

Many HUGIN API functions that operate on nodes also modify the state
of the domain or class to which the nodes belong. For example, entering
evidence to a node clearly modifies the state of the node, but it also modifies
book-keeping information relating to evidence within the domain to which
the node belongs.
On the other hand, many HUGIN API functions only read attributes of a
class, domain, or node. Such functions can be used simultaneously from
different threads on the same or related objects, as long as it has been ensured that no thread is trying to modify the objects concurrently with the
read operations. Examples of functions that only read attributes are: h
node get category(30) , h domain get attribute(46) , h node get belief (125) , etc.
In general, all functions with get or is as part of their names do not
modify data, unless their descriptions explicitly state that they do. Examples
of the latter category are:
h node get name(42) and h class get name(53) will assign names to the
node or class, if no name has previously been assigned. (If the node
or class is known to be named, then these functions will not modify
data.)
h node get table(39) , h node get experience table(160) , and h node get
fading table(161) will create a table if one doesnt already exist.
h domain get marginal(126) and h node get distribution(126) must, in
the general case, perform a propagation (which needs to modify the
junction tree).
25

 All HUGIN API functions returning a list of nodes may have to allocate
and store the list.

26

Chapter 2

Nodes and Domains

The functions described in this chapter allow an application to construct and
modify flat belief network and LIMID models, known as domains. Chapter 3 provides functions for constructing object-oriented models for belief
networks and LIMIDs. An object-oriented model must be converted to a
domain before it can be used for inference.
In HUGIN, dynamic models (known as Dynamic Bayesian Networks) must be
constructed as object-oriented models and converted to domains for inference using special functions. See Chapter 4 for further information.
A large part of the functions (those that operate on nodes) described in this
chapter can also be used for nodes in object-oriented (including dynamic)
models. If special conditions apply when a function is used for such nodes,
they are stated in the description of the function, prefixed with OOBN:,1
DBN:, or OOBN/DBN:.

2.1

Types

Nodes and domains are the fundamental objects used in the construction of
belief network and LIMID models in HUGIN. The HUGIN API introduces the
opaque pointer types h node t and h domain t to represent these objects.

2.1.1

Node category

In ordinary belief networks, all nodes represent random variables. But, in

LIMIDs, nodes also represent decisions and utility functions. And in objectoriented models, nodes also represent class instances see Section 3.8.
1
Although slightly misleading, we use the abbreviation OOBN to refer to object-oriented
belief networks as well as object-oriented LIMIDs.

27

In all of these network models, so-called function nodes [26] can be created:
A function node represents either a single real value (a real-valued function
node) or a discrete marginal distribution (a discrete function node). In both
cases, this entity is a function of (the values or distributions of) the parents.
Real-valued function nodes are not (directly) involved in the inference process evidence cannot be specified for such nodes, but the functions associated with the nodes can be evaluated using the results of inference or
simulation as input. In this case, evaluation must take place after inference
or simulation.
However, real-valued function nodes can also be used to provide input to
the table generation process: The expressions in the model of a node may
refer to the values of (real-valued) function nodes in the parent set of the
node. Since table generation takes place before inference, the values of
these parents must be available before inference.
A discrete function node behaves like a chance node, except that the parents
of the node are not affected by inference with evidence specified for the
node. Instead, the values or distributions of the parents are used to generate
a marginal distribution (which becomes the node table) for the node before
inference is performed.
These properties allow the tables of some nodes to depend on the beliefs
(computed by inference) of other nodes see Section 10.2. In order to
make this work, the network must satisfy some constraints see Section 2.4.
In order to distinguish between the different types of nodes, the HUGIN API
associates with each node a category, represented as a value of the enumeration type h node category t. The constants of this enumeration type are:
h category chance (for nodes representing random variables),
h category decision (for nodes representing decisions),
h category utility (for nodes representing utility functions),
h category function (for function nodes), and
h category instance (for nodes representing class instances in objectoriented models).
In addition, the special constant h category error is used for handling errors.

2.1.2

Node kind

Another grouping of nodes exists, called the kind2 of a node. This grouping is a characterization of the state space of the node. The HUGIN API
introduces the enumeration type h node kind t to represent it.
2
The terms category and kind have been deliberately chosen so as not to conflict with the
traditional vocabulary used in programming languages. Thus, the term type was ruled out.

28

Chance and decision nodes are either discrete or continuous.3 The enumeration constants h kind discrete and h kind continuous represent those
kinds. Discrete nodes have a finite number of states. Continuous nodes
are real-valued and have a special kind of distribution, known as a Conditional Gaussian (CG) distribution, meaning that the distribution is Gaussian
(also known as normal) given values of the parents. For this reason, continuous nodes are also referred to as CG nodes. (See Appendix A for further
information on CG variables.)
As mentioned above, function nodes are either real-valued or discrete. Discrete function nodes have discrete kind, while real-valued function nodes
have other kind (represented by the enumeration constant h kind other).
This is also the kind of utility and instance nodes.
In addition, the special constant h kind error is used for handling errors.

2.2

Domains: Creation and deletion

A domain is used to hold all information associated with a (non-OOBN)

network model.
h domain t h new domain (void)

Create a new empty domain. If creation fails, NULL is returned.

When a domain is no longer needed, the internal memory used by the domain can be reclaimed and made available for other purposes.
h status t h domain delete (h domain t domain)

Release all (internal) memory resources used by domain.

All existing references to objects owned by domain (for example, nodes) are
invalidated by this operation, since those objects are also deleted.
A domain can also be created by cloning an existing domain.
h domain t h domain clone (h domain t domain)

Create a clone of domain. The clone will be identical to domain, except that
the clone will not be compiled (even if domain is compiled).4
3

Currently, the HUGIN API does not support LIMIDs with continuous nodes. Thus, all
(chance and decision) nodes of a LIMID must be discrete.
4
Chapter 7 provides information on compiling domains a prerequisite for performing
inference.

29

2.3

Nodes: Creation and deletion

The following function is used for creating nodes in a domain. Only chance,
decision, utility, and function nodes are permitted in a domain.
x

h node t h domain new node

(h domain t domain, h node category t category,
h node kind t kind)
Create a new node of the specified category and kind within domain. The
new node has default (or no) values assigned to its attributes: for example,
it has no name, it has just one state (if the node is a discrete node), and it
has no table. The attributes of the new node must be explicitly set using the
relevant API functions.
If the new node is not a real-valued function node, and domain is compiled,
the corresponding compiled structure is removed since it no longer reflects
the domain (see Section 7.5).
If an error occurs, the function returns NULL.

h domain t h node get domain (h node t node)

Retrieve the domain to which node belongs. If node is NULL, or it belongs to
a class, NULL is returned.

h node category t h node get category (h node t node)

Return the category of node. If node is NULL, h category error is returned.

h node kind t h node get kind (h node t node)

Return the kind of node. If node is NULL, h kind error is returned.
In influence diagrams, it is sometimes convenient to be able to convert a
decision node to a chance node (or vice versa).

h status t h node set category

(h node t node, h node category t category)
Set the category of node to category node must be a discrete (chance or
decision) node, and category must be h category chance or h category decision.
OOBN/DBN: node must not be an output clone or a temporal clone.
Unless category is equal to the current category of node, and if node belongs
to a compiled domain, then that domain is uncompiled (see Section 7.5).
30

The following function is intended for modifying a network. If a complete

domain is to be disposed off, use h domain delete(29) instead.
h status t h node delete (h node t node)

Delete node (and all links involving node) from the domain or class to which
node belongs. If node has children, the tables and models of those children
are adjusted (see h node remove parent(34) for a description of the adjustment procedure).
If node is not a real-valued function node, and it belongs to a domain, then
that domain is uncompiled (see Section 7.5).
OOBN: Special actions are taken if node is an interface node, an instance
node, or an output clone. See Section 3.7 and Section 3.8 for further details.
DBN: If node has a temporal clone, then that clone is also deleted.
A new node can also be created by cloning an existing node.
h node t h node clone (h node t node)

Create a clone of node. The clone belongs to the same domain or class
as node, and it has attributes that are identical to (or clones of) the corresponding attributes of node: category, kind, subtype, number of states, state
labels, state values, parents, tables (conditional probability, policy, utility,
experience, and fading), model, case data, evidence, structure learning constraints, label, and user-defined attributes. However, the user data pointer
is not copied (it is set to NULL for the clone).
The clone has no name (because there cannot be two identically named
nodes in the same domain or class). Also, the clone has no children (because
that would imply changes to the children).
OOBN/DBN: If node is an interface node or an output clone, or node is or
has a temporal clone, then the clone has none of these properties.
OOBN: If node is a class instance, then the clone is an instance of the same
class as node and has the same inputs as node.
OOBN: If node belongs to a domain derived from a class (a so-called runtime domain), then the source list of node is not copied to the clone.
If (and only if) the cloning process succeeds, node is not a real-valued function node, and it belongs to a domain, then that domain is uncompiled
(see Section 7.5).

2.4

The links of the network

The links of a belief network or a LIMID are directed edges between the
nodes of the network. [Undirected edges are also possible, but the API
interface to support them has not yet been defined. However, see Chapter 13
for a description of the NET language interface.]
31

If there exists a directed edge from a node u to a node v, we say that u is

a parent of v and that v is a child of u. The HUGIN API provides functions
for adding and removing parents to/from a node, replacing a parent with
another compatible parent, reversing a directed edge between two nodes,
as well as functions for retrieving the current set of parents and children of
a node.
The semantics of links depend on the categories of the nodes involved. For
chance, decision, and utility nodes, real-valued function parents are used to
provide input to the table generation process: The expressions in the model
of the child node may refer to the values of function parents. For all other
types of parents, the semantics of links vary by the different categories of
the child node:
For chance nodes, the incoming links represent probabilistic dependence: The distribution of a chance node is conditionally independent
of all non-descendants of the node given its (non-real-valued function)
parents.
For decision nodes, the incoming links represent availability of information: The parents of a decision node are assumed to be known
when the decision is made.
For utility nodes, the parents represent the set of nodes that the utility
function depends on.
For real-valued function nodes, the parents represent the set of nodes that
the function associated with the node is allowed to depend on.
For discrete function nodes, all parents are used to provide input to the table
generation process (like real-valued function parents do for non-function
nodes).
Chance, decision, and utility nodes can only have chance, decision, and
function nodes as parents. Nodes representing class instances must have no
parents and no children.
The network cannot be an arbitrary directed graph: It must be acyclic. Moreover, in order to prevent cyclic dependencies in the evaluation of function
nodes, additional constraints must be imposed on the topology of the network. These constraints can be formulated as follows.
A link from u to v (u v) is called a functional link if (and only if) u is
a real-valued function node or v is a function node (of any kind). A trail
from u to v is a sequence of nodes u = x1 , x2 . . . , xn = v such that either
xi xi+1 or xi xi+1 and xi xi+1 is not a functional link (1 i < n).
A trail that contains a functional link is called a functional trail, and a trail
from u to u is called a cyclic trail.
The network must not contain cyclic functional trails.
32

Moreover, the SPU (Single Policy Updating) algorithm [24] used to compute policies in LIMIDs cannot handle networks with functional trails from
a node u to a node v unless u or v is a real-valued function node. An attempt to create a LIMID that violates this condition results in a functional
dependency error.5
It is not possible to link nodes from different domains or classes.
The quantitative part of the relationship between a discrete, a continuous,
or a utility node and its parents (ignoring functional links) is represented as
a table. This table can either be specified directly (as a complete set of numbers), or it can be generated from a model. For real-valued function nodes,
only models can be used. When links are added, removed, or reversed, the
tables and models involved are automatically updated.
x

h status t h node add parent (h node t child, h node t parent)

Add parent as a new parent of child (i.e., add a directed link from parent to
child). If child is a chance, a decision, or a utility node, then parent must be
a chance, a decision, or a function node. And if child is a discrete (chance
or decision) node, then parent must be a discrete node or a function node.
OOBN: child and parent must not be class instances, and child must not be
an input node or an output clone.
DBN: If child is a temporal clone, then parent must also be a temporal clone.
If adding the link would create a directed cycle in the network, the operation
is not performed. The operation is also not performed if the link would cause
the constraints described above concerning functional trails to be violated.
If parent is discrete (and child is not a function node), the tables6 (if any)
of child are updated as follows: The value associated with a given state
configuration of the new table is the same as the value associated with the
corresponding state configuration (that is, exclusive the value of parent) of
the old table. If child and parent are both continuous, the (i)-parameters
(see h node set beta(41) ) associated with parent in the updated conditional
probability table of child are set to zero. In all other cases, the tables of child
do not need to be updated.
The model (if any) of child is not affected by this operation.
Finally, if the new link is a non-functional link, and the nodes belong to a
domain, then that domain will be uncompiled (see Section 7.5).
5
Also notice that some HUGIN API operations do not support (non-LIMID) networks with
such functional trails. The operations are simulation, value-of-information analysis, sensitivity analysis, finding most probable configurations, and the EM algorithm.
6
In addition to a conditional probability table, two other tables can be associated with
a chance node: An experience and a fading table (see Section 11.1) can be created for the
purpose of parameter learning. All tables are updated in the same manner.

33

h status t h node remove parent (h node t node, h node t parent)

Delete the directed link from parent to node. The tables of node are updated
as follows: If parent is discrete (and node is not a function node), the contents of the updated table are the parts of the old table corresponding to
state 0 of parent (see Section 2.5). If node and parent are both continuous,
the (i)-parameters (see h node set beta(41) ) corresponding to the parent
node link are deleted from the table. In all other cases, the tables of node do
not need to be updated.
The model (if any) of node is updated as follows: If parent is a model node,
expressions corresponding to state 0 of parent are kept (all other expressions
are deleted). In any case, all expressions that refer to parent are deleted.
Finally, if the parent node link is a non-functional link, and the nodes belong to a domain, then that domain will be uncompiled (see Section 7.5).

h status t h node switch parent

(h node t node, h node t old parent, h node t new parent)
Substitute new parent for old parent as a parent of node, while preserving
the validity of the tables and model of node (all references to old parent are
replaced by references to new parent). The old parent and new parent nodes
must be compatible see below for the definition of compatibility.
DBN: If node is a temporal clone, then new parent must also be a temporal
clone.
If switching parents would create a directed cycle in the network, or if it
would cause the constraints described above concerning functional trails to
be violated, the operation is not performed.
As usual, if the affected links are non-functional links, and the nodes belong
to a domain, then that domain will be uncompiled (see Section 7.5).
In order for two nodes to be compatible, the following conditions must hold:
The nodes must have
the same category and kind;
the same subtype and the same number of states (if the nodes are
discrete);
the same list of state labels (if the nodes are labeled);
the same list of state values (if the nodes are numbered or of interval
subtype).
The motivation for this definition is that compatible nodes should be interchangeable with respect to the table generator (see Chapter 6). That is,
replacing one node in a model with another compatible node should not
34

affect the table produced by the table generator.7 That is also the reason
for requiring the lists of state labels to be identical only for labeled nodes,
although all discrete nodes can have state labels.
x

h status t h node reverse edge (h node t node1 , h node t node2 )

Reverse the directed edge between node1 and node2 (which must be chance
nodes of the same kind). This is done such that the joint probability distribution defined by the modified network is equivalent to the joint probability
distribution defined by the original network. In order to accomplish this,
node1 inherits the parents of node2 (except node1 , of course), and vice versa
for node2 .
OOBN: The nodes must not be input nodes or output clones.
DBN: The nodes must be of the same temporal order (that is, both or none
of the nodes must be temporal clones).
The operation is not performed, if reversal of the edge would create a directed cycle in the network.
The experience and fading tables (see Section 11.1) as well as models (if
any) of node1 and node2 are deleted.
Finally, if the nodes belong to a domain, then that domain will be uncompiled (see Section 7.5).

h node t h node get parents (h node t node)

Return the parents of node (as a NULL-terminated list). If an error occurs,
NULL is returned.
The list of parents is stored within the node data structure. The application
must not modify or deallocate this list.
When a new discrete parent is added to node, the parent is added at the
front of the list of parents. A new parent of any other type is added at the
end of the list.
Example 2.1 The following code prints the names of the parents of a node:
h_node_t n, *parents, *p;
...
if ((parents = h_node_get_parents (n)) == 0)
/* handle error */;
else
{
printf ("Parents of %s:\n", h_node_get_name (n));
for (p = parents; *p != 0; p++)
printf ("%s\n", h_node_get_name (*p));
}
7
But note that this does not hold for real-valued function parents (introduced in HUGIN
API 7.6) see Section 6.8.

35

If the list returned by h node get parents (or h node get children) is used to
control the iteration of a loop (such as the for-loop in the example above),
then API functions that modify the list must not be used in the body of the
loop. For example, calling h node add parent(33) modifies the list of parents
of the child and the list of children of the parent: The contents of the lists are
obviously modified, but the memory locations of the lists might also change.
Other similar functions to watch out for are h node remove parent(34) , h
node switch parent(34) , h node reverse edge(35) , and h node delete(31) .
The problem can be avoided if a copy of the list is used to control the loop.
h node t h node get children (h node t node)

Return the children of node (as a NULL-terminated list). If an error occurs,

NULL is returned.
The list of children is stored within the node data structure. The application
must not modify or deallocate this list.

2.4.1

The requisite parents and ancestors of decision nodes

Not all available observations matter when a decision must be made. Intuitively, a parent of a decision node is said to be requisite if the value of the
parent may affect the optimal choice of the decision.
The performance of inference in a LIMID can be improved, if the network is
simplified by removing the nonrequisite parents of all decision nodes.
Lauritzen and Nilsson [24] present an algorithm for removing the nonrequisite parents of the decision nodes in a LIMID. The result of this algorithm is
a LIMID, where all parents of each decision node are requisite. This network
is known as the minimal reduction of the LIMID.8
h node t h node get requisite parents (h node t node)

Return a list of the requisite parents of node (which must be a decision node
belonging to a domain). If an error occurs, NULL is returned.
If the requisite parents are not already known (see below), the function
computes the minimal reduction of the underlying LIMID network. The
parents of node in this minimal reduction are returned.
Notice that the function does not remove the nonrequisite parents it only
tells which parents are (non)requisite. In order to remove the nonrequisite
parents, h node remove parent(34) must be used.
In order to improve performance, the results of the minimal reduction algorithm are cached (that is, the requisite parents of all decisions are cached).
8

Our definition of requisiteness is slightly different than the one given by Lauritzen and
Nilsson: We define a parent of a decision node to be requisite if and only if it is also a parent
of the decision node in the minimal reduction of the LIMID.

36

If a link is added, if a link is removed (unless that link represents a nonrequisite parent of a decision node), or if a decision node is converted to
a chance node (or vice versa), the cached results are deleted. Notice that
deletion of nodes usually removes links (but creation of nodes does not add
links). Also notice that switching parents adds a link and removes a link and
therefore also causes the cached results to be deleted.
The list of requisite parents is stored within the node data structure. The
application must not modify or deallocate this list. Also, this list is deleted
by the HUGIN API when the contents become invalid (as explained above).
In this case, an updated list must be requested using h node get requisite
parents.
In ordinary influence diagrams, decisions are made according to the noforgetting rule (which states that past observations and decisions are taken
into account by all future decisions). To help identify the relevant part of
the past, the following function can be used.
h node t h node get requisite ancestors (h node t node)

Return a list of the requisite ancestors of node (which must be a decision

node belonging to a domain). If an error occurs, NULL is returned.
The requisite ancestors are found by augmenting the LIMID as follows:
If there is a directed path from a decision D to another decision D 0 , then
D and its parents are added as parents of D. Then the minimal reduction
algorithm is applied to the augmented LIMID. The resulting set of parents
of a decision in the reduced LIMID is the requisite ancestors of the decision.
Notice that the function does not modify the network. In order to obtain
the influence diagram solution of the decision problem (i.e., the solution
implied by the no-forgetting rule), the requisite ancestors must be added
as parents of the decision nodes in the LIMID before the SPU procedure is
invoked.9
Like the h node get requisite parents function, the h node get requisite ancestors function also employs a caching scheme to improve performance. If
no cached results are available, the minimal reduction algorithm is applied
to an augmented LIMID as described above. Subsequent calls then use the
cached results.
Certain operations cause the cached results to be deleted: If a link is removed, if a link is added (unless that link represents a requisite ancestor of
a decision node), or if a decision node is converted to a chance node (or
vice versa), the cached results are deleted.
The list of requisite ancestors is stored within the node data structure. The
application must not modify or deallocate this list. Also, this list is deleted
9

Ordinary influence diagrams also assume a total ordering of all decisions in the network.
This ordering can be specified by ensuring the existence of a directed path that contains all
decisions in the network. This must be done before the requisite ancestors are retrieved.

37

by the HUGIN API when the contents become invalid (as explained above).
In this case, an updated list must be requested using h node get requisite
ancestors.

2.5

The number of states of a discrete node

As mentioned above, discrete nodes in the HUGIN API has a finite number
of states. The enumeration of the states follows traditional C conventions: If
a node has n states, the first state has index 0, the second state has index 1,
. . . , and the last state has index n 1.
It is possible to associate labels and values with the states of discrete nodes.
See Section 6.5 and Section 6.6.
The following function is used to specify the number of states of a discrete
node.
x

h status t h node set number of states (h node t node, size t count)

Set the number of states of node to count (node must be a discrete node,
and count must be a positive integer). Unless count is equal to the current
number of states of node, any evidence specified for node (see Section 9.1.1)
is removed, and if node belongs to a domain then that domain will be uncompiled (see Section 7.5).
OOBN/DBN: node must not be an output clone or a temporal clone.
If node is a boolean node, then count must be 2 (that is, a boolean node must
always have two states).
Changing the number of states of a node has implications for all tables in
which the node appears. The affected tables are the table associated with
the node itself and the tables associated with the (non-function) children
of the node. (See Section 2.6 and Section 11.1 for more information about
those tables.)
Let hN1 , . . . , Nk , . . . , Nl i be the node list of some table for which the number
of states of node Nk is being changed from nk to mk , and let hi1 , . . . , ik1 ,
ik , ik+1 , . . . , il i be a configuration of that table (see Section 5.1 for an explanation of node lists and configurations). If mk < nk , the updated table is
obtained by deleting the data associated with the configurations for which
ik mk . If mk > nk , the updated table is obtained by copying the data associated with the configuration hi1 , . . . , ik1 , 0, ik+1 , . . . , il i to the new configurations (i.e., the configurations for which ik nk ).
Changing the number of states of a node might also have implications for the
models of the children of the node. If the node appears as a model node
in the model of a child, then the model is resized as follows: Expressions
corresponding to deleted configurations are deleted, and expressions corresponding to new configurations are set to NULL.
38

If count is smaller than the current number of states of node, then the labels
and the values associated with the deleted states are deleted.
OOBN: If node is an output node of its class, then the changes described
above are applied to all its output clones (recursively, if output clones are
themselves output nodes).
DBN: If node has a temporal clone, then the changes described above are
applied to that clone.
h count t h node get number of states (h node t node)

Return the number of states of node (which must be a discrete node). If an

error occurs, a negative number is returned.

2.6

The table of a node

A table is associated with each (discrete, continuous, and utility) node10

in a belief network or a LIMID model. This table has different semantics
depending on the category of the node:
For a chance node, the table is a conditional probability table (CPT).
For a decision node, the table represents a policy that prescribes what
the decision maker should do given observations of the parents. This
table is similar to a CPT (i.e., the table contains conditional probability
distributions, but the distributions are usually deterministic).
For a discrete function node, the table is a marginal probability table
(i.e., the parents are not included in the table). This table is usually
generated from a model as this is the only way the parents can influence the child node.
For a utility node, the table represents a utility function.
OOBN: Nodes representing class instances do not have tables. Also, output
clones do not have tables.
For utility and discrete nodes, the contents of the tables can be specified
directly (by writing to the data arrays of the tables) or indirectly (using the
table generator see Chapter 6). For continuous nodes, the parameters of
the CG distributions must be specified using appropriate functions.
x

h table t h node get table (h node t node)

Retrieve the table associated with node (which must be a discrete, a continuous, or a utility node). If an error occurs, NULL is returned.
10
Instead of tables, real-valued function nodes use models to express the (functional) relationship between the child and its parents. See Chapter 6.

39

OOBN: node must not be a class instance or an output clone.

The returned table is the real thing (i.e., not a copy), and the contents of
the table can be modified using functions that provide access to the internal
data structures of tables (see Chapter 5).
The node list of the table contains the discrete parents of node (unless node is
a function node), followed by node (unless node is a utility node), followed
by the continuous parents of node (unless node is a function node).11 The
order of the parents in this node list is the same as the order in the node list
returned by h node get parents(35) , unless the order has been changed
see h table reorder nodes(74) . (The node list determines the configurations
of the table see Section 5.1.)
The table is not created until it is needed (for example, requested by this
function or required in order to compile a domain). It is recommended to
specify the parents of a node before specifying its table as this simplifies the
creation of links (because then the table wont have to be resized for each
link being added).
However, if the table has been created, and the set of parents of node is
changed, or the number of states of node or one of its parents is changed
(assuming discrete nodes), then the table is resized accordingly. Because of
that, any pointers to the internal data structures of the table (for example,
a pointer to the data array of the table) that the application may be holding
become invalid and must be retrieved again from the table.
If node is a discrete node, the entries of the table corresponding to a given
parent state configuration should be nonnegative and sum to 1. If not, the
inference engine (that is, the propagation operation) will adjust (normalize) the table entries in order to satisfy this condition. There are no such
restrictions on utility tables.
It is also possible to specify the contents of a node table indirectly through
the table generation facility (see Chapter 6). If this facility is used for some
node table, then the contents of that table is generated from a mathematical description of the relationship between the node and its parents. It is
possible to modify the contents generated from such a description, but note
that the inference engine will regenerate the table when certain parameters
are changed (see Section 6.8 for precise details).
If the contents of a table is changed, the updated table will be used by the
inference engine, provided it has been notified of the change. HUGIN API
functions that change node tables automatically provide this notification.
However, for changes made by storing directly into a table (i.e., using the
array pointer returned by h table get data(72) for storing values), an explicit
notification must be provided. The following function does this.
11
Notice that real-valued function parents of node (if any) are not represented in the table.
Also notice that the table of a discrete function node only contains the node itself.

40

h status t h node touch table (h node t node)

Notify the inference engine that the table of node has been modified by
storing directly into its data array. This notification must be provided before
subsequent calls to other HUGIN API functions.
Omission of such a notification may cause the inference engine to malfunction!
Example 2.2 This piece of code shows how to specify the probability table for the
variable A in the Chest Clinic belief network [25]. This variable is binary and has
no parents: P(A = yes) = 0.01 and P(A = no) = 0.99.
h_node_t A;
h_table_t table = h_node_get_table (A);
h_number_t *data = h_table_get_data (table);
data[0] = 0.01;
data[1] = 0.99;
h_node_touch_table (A);

The conditional distribution for a continuous random variable Y with discrete parents I and continuous parents Z is a (one-dimensional) Gaussian
distribution conditional on the values of the parents:
p(Y |I = i, Z = z) = N((i) + (i)T z, (i))
[This is known as a CG distribution.] Note that the mean depends linearly
on the continuous parent variables and that the variance does not depend
on the continuous parent variables. However, both the linear function and
the variance are allowed to depend on the discrete parent variables. (These
restrictions ensure that exact inference is possible.)
The following six functions are used to set and get the individual elements
of the conditional distribution for a continuous node. In the prototypes of
these functions, node is a continuous chance node, parent is a continuous
parent of node, i is the index of a discrete parent state configuration12 (see
Section 5.1 for an explanation of configuration indexes), and alpha, beta,
and gamma refer to the (i), (i), and (i) components of a CG distribution
as specified above.
x

h status t h node set alpha

(h node t node, h double t alpha, size t i)

h status t h node set beta

(h node t node, h double t beta, h node t parent, size t i)
12
If node has no discrete parents, then there is only one configuration the empty configuration. In this case, i must be 0.

41

h status t h node set gamma

(h node t node, h double t gamma, size t i)

Here, gamma must be nonnegative.

x

h double t h node get alpha (h node t node, size t i)

h double t h node get beta (h node t node, h node t parent, size t i)

h double t h node get gamma (h node t node, size t i)

For the last three functions: If an error is detected, a negative value is returned, but this is not of any use for error detection (except for h node get
gamma), since any real value is valid for both the (i) and (i) parameters.
Thus, errors must be checked for using the h error code(18) function.

2.7

The name of a node

The HUGIN system uses a number of text files:

NET files for storing specifications of networks,
data files for storing learning data,
case files for storing a single case,
node list files (e.g., for storing elimination orders for triangulations),
log files for generating logs of actions performed by some HUGIN API
calls (e.g., compilation operations).
In order to refer to a node in text written to any of these files, the node
must be named. If the node doesnt have a name, a name is automatically
assigned as part of the operation that generates the text. But nodes can also
be explicitly named using the following function.
x

h status t h node set name (h node t node, h string t name)

Create a copy of name and assign it to node. No other node in the network
(a domain or a class) to which node belongs must have the same name. The
name must have the same form as a C identifier, except that if node belongs
to a domain, then the name can be a sequence of identifiers, separated by
single dots (periods) see Section 13.8. The reason for this exception is to
permit naming (using meaningful names) of nodes of a runtime domain.

h string t h node get name (h node t node)

Retrieve the name of node. If node has not previously been assigned a name,
a valid name is automatically assigned. If an error occurs, NULL is returned.
42

Note that the name returned by h node get name is not a copy. Thus, the
application must not modify or free it.
h node t h domain get node by name
(h domain t domain, h string t name)

Return the node with name name in domain, or NULL if no node with that
name exists in domain, or if an error occurs (i.e., domain or name is NULL).

2.8

Iterating through the nodes of a domain

An application may need to perform some action for all nodes of a domain.
To handle such situations, the HUGIN API provides a set of functions for
iterating through the nodes of a domain, using an order determined by the
age of the nodes: the first node in the order is the youngest node (i.e., the
most recently created node that hasnt been deleted), . . . , and the last node
is the oldest node.
If the application needs the nodes in some other order, it must obtain all the
nodes, using the functions described below, and sort the nodes according to
the desired order.
x

h node t h domain get first node (h domain t domain)

Return the first node of domain, using the order described above, or NULL if
domain contains no nodes, or domain is NULL (this is considered an error).

h node t h node get next (h node t node)

Return the node that follows node in the order, or NULL if node is the last
node in the order, or if node is NULL (which is considered an error).
Example 2.3 This function counts the nodes of a given domain.
int count_nodes (h_domain_t d)
{
h_node_t n;
int count = 0;
for (n = h_domain_get_first_node (d); n != 0;
n = h_node_get_next (n))
count++;
return count;
}

43

2.9

User data

Applications sometimes need to associate data with the nodes of a domain

(or the domain itself). Examples of such data: the window used to display
the beliefs of a node, the last time the display was updated, the external
source used to obtain findings for a node, etc.
The HUGIN API provides two ways to associate user data with domains and
nodes:
as arbitrary data, managed by the user, or
as attributes (key/value pairs where the key is an identifier, and the
value is a character string), managed by the HUGIN API.

2.9.1

Arbitrary user data

The HUGIN API provides a slot within the node structure for exclusive use
by the application. This slot can be used to hold a pointer to arbitrary data,
completely controlled by the user.
x

h status t h node set user data (h node t node, void p)

Store the pointer p in the user data slot of node.

void h node get user data (h node t node)

Return the value stored in the user data slot of node. If no value has been
stored, the stored value is NULL, or node is NULL (this is an error), NULL is
returned.
No other functions in the HUGIN API touch the user data slots of the nodes.
Example 2.4 In an application displaying the beliefs of nodes in windows, each
node will have a window associated with it. The simplest way to keep track of
these belief windows is to store them in the user data fields of the nodes.
Creating and storing a belief window can be done in the following way:
belief_window w;
h_node_t n;
...
w = create_belief_window (n);
h_node_set_user_data (n, (void*) w);
where create belief window is a function defined by the application. Note the cast
to type void in the call to h node set user data (for this to work properly, belief
window should be a pointer type).
Now, the belief window can be used in the following way:

44

belief_window w;
h_node_t n;
...
w = (belief_window) h_node_get_user_data (n);
update_belief_window (w, n);
where update belief window is a function defined by the application. Again, note
the cast of the pointer type.

Using the user data facility is analogous to adding an extra slot to the node
data structure. It must be noted that only one such slot is available. If more
are needed, store a list of slots or create a compound data type (e.g., a C
structure). Note also that the extra slot is not saved in HUGIN KB or in
NET files. If this is needed, the application must create the necessary files.
Alternatively, the attribute facility described below can be used.
It is also possible to associate user data with a domain as a whole. This is
done using the following functions.
x

h status t h domain set user data (h domain t domain, void p)

Store the pointer p in the user data slot of domain.

void h domain get user data (h domain t domain)

Return the value stored in the user data slot of domain. If no value has been
stored, the stored value is NULL, or domain is NULL (this is an error), NULL is
returned.

2.9.2

User-defined attributes

In the previous section, we described a way to associate arbitrary data with a

node or a domain object. That data can be anything (for example, a list or a
tree). However, if more than one data object is required, the user must build
and maintain a data structure for the objects himself. Moreover, the data are
not saved in the HUGIN KB (Section 2.10) and the NET files (Chapter 13).
Sometimes we need the ability to associate several user-specified data objects with domains and nodes and to have these data objects saved in the
HUGIN KB and the NET files. The HUGIN API provides this feature but at the
cost of requiring the data to be C-strings (i.e., sequences of non-null characters terminated by a null character). Each data object (string) is stored
under a name a key, which must be a C-identifier (i.e., a sequence of letters and digits, starting with a letter, and with an underscore counting as a
letter).
The HUGIN GUI tool stores strings using the UTF-8 encoding. If you intend
to have your networks loaded within that tool, you should make sure to
store your attributes using the UTF-8 encoding.
45

h status t h node set attribute

(h node t node, h string t key, h string t value)
Insert (or update, if key is already defined) the key/value-pair in the attribute
list for node. If value is NULL, the attribute is removed.
OOBN: If node is an output clone, then the attribute is not saved if the class
is saved as a NET file (because the NET file format doesnt support that).

h string t h node get attribute (h node t node, h string t key)

Lookup the value associated with key in the attribute list for node. If key is
not present, or if an error occurs, NULL is returned.
The string returned by h node get attribute is stored in the attribute list for
node. The application must not deallocate this string.
The following two functions perform similar operations on domains.

h status t h domain set attribute

(h domain t domain, h string t key, h string t value)

h string t h domain get attribute

(h domain t domain, h string t key)
If you want to create your own attributes, pick some attribute names that
are not likely to clash with somebody elses choices (or with names that the
HUGIN API or the HUGIN GUI tool might use in the future). For example,
use a common prefix for your attribute names.
In order to access the values of attributes, one must know the names of the
attributes in advance. The following set of functions provides a mechanism
for iterating over the list of attributes associated with a given node or domain.
The notion of an attribute as a key/value pair is represented by the opaque
pointer type h attribute t.

h attribute t h node get first attribute (h node t node)

Retrieve the first attribute object for node. If the attribute list is empty (or
node is NULL), NULL is returned.

h attribute t h domain get first attribute (h domain t domain)

Retrieve the first attribute object for domain. If the attribute list is empty (or
domain is NULL), NULL is returned.
The attributes returned by these functions are actual objects within the attribute lists of node or domain. Do not attempt to deallocate these objects.
46

h attribute t h attribute get next (h attribute t attribute)

Retrieve the attribute object that follows attribute in the attribute list containing attribute. If attribute is the last object in the list, or attribute is NULL
(this is a usage error), NULL is returned.
Given an attribute object, the following functions can be used to retrieve the
key and value parts of the attribute.

h string t h attribute get key (h attribute t attribute)

h string t h attribute get value (h attribute t attribute)

Retrieve the key or value associated with attribute. (These are the actual
strings stored within the attribute do not modify or deallocate them.)
Note that h node set attribute(46) and h domain set attribute(46) modify (or
even delete, if the value argument is NULL) objects in the attribute lists for
the affected node or domain. If the application is holding (a pointer to) an
attribute, and then calls one of these functions to change the value of (or
even delete) the attribute, then (pointers to) the old value (or the attribute
itself, if the value argument was NULL) are no longer valid. These facts
should be kept in mind when iterating over attribute lists.

2.10

HUGIN Knowledge Base files

When a domain has been created, it can be saved to a file in a portable

binary format. Such a file is known as a HUGIN Knowledge Base (HUGIN KB,
or simply HKB, for short) file. By convention, we name such files using the
extension .hkb. (A domain can also be saved in a textual format, called the
NET format see Chapter 13.)
When a domain is loaded from an HKB file, the state of the domain is
generally the same as when it was saved with the following exceptions:
The case data used by the structure learning and EM algorithms (see
Chapter 12) are not stored in the HKB file.
If the domain is compiled (but not compressed), then the contents of
the HKB file include only information about the structure of the junction trees, not the numerical data stored in the tables of the junction
trees. In other words, a compiled domain is saved as if it was only
triangulated. When the domain is loaded from the HKB file, it must
be explicitly compiled (using h domain compile(103) ) before it can be
used for inference13 this will use the already constructed junction
trees.
13
This is a change in HUGIN API version 6.6. In previous versions of the HUGIN API,
the compilation was attempted as part of the load operation except in very old versions,

47

 If the domain is compressed (which implies compiled), then the domain will also be compressed when loaded. This property implies that
compressed domains can be created on computers with large amounts
of (virtual) memory and then later be loaded on computers with limited amounts of (virtual) memory.
When a compressed domain is loaded, a propagation is required before beliefs can be retrieved.14
If the domain is a triangulated influence diagram (i.e., the HKB file
was created by a HUGIN API older than version 7.0), then the domain
is loaded in uncompiled form. The domain is then treated as a LIMID.
OOBN/DBN: The source lists of nodes in a runtime domain are not
saved in the HKB file. This implies that the domain is not recognized
as a runtime domain when it is loaded from the HKB file.15
There is no published specification of the HKB format, and since the format
is binary (and non-obvious), the only way to load an HKB file is to use the
appropriate HUGIN API function. This property makes it possible to protect
HKB files from unauthorized access: A password can be embedded in the
HKB file, when the file is created; this password must then be supplied,
when the HKB file is loaded. (The password is embedded in the HKB file in
encrypted form, so that the true password cannot easily be discovered by
inspection of the HKB file contents.)
In general, the format of an HKB file is specific to the version of the HUGIN
API that was used to create it. Thus, when upgrading the HUGIN API (which
is also used by the HUGIN GUI tool, so upgrading that tool usually implies
a HUGIN API upgrade), it may be necessary to save a domain in the NET
format (see Section 13.10) using the old software before upgrading to the
new version of the software (because the new software may not be able to
load the old HKB files).16
where the table data were actually included in the HKB file, and therefore the table data
could be loaded directly from the file.
14
This is also a change in HUGIN API version 6.6. See the previous footnote (replacing
compilation with propagation).
15
This prevents functions such as h domain learn class tables(179) from being used.
16
The HKB formats for HUGIN API versions 3.x and 4.x were identical, but the HKB format
changed for version 5.0 and again for versions 5.1, 5.2, and 5.3. Versions 5.4, 6.0, and 6.1
used the same format as version 5.3. Versions 6.26.5 also used this format for HKB files
that were not password protected, but a newer revision of the format was used for password
protected HKB files. The HKB format changed for versions 6.6, 6.7, 7.0, and 7.1. Version 7.2
used the same format as version 7.1. Version 7.3 also used this format for networks without
function nodes, but a newer revision of the format was used for networks with function
nodes. Versions 7.4 and 7.5 used the same format as version 7.3, unless the file contained a
compressed domain in this case, a newer revision of the format was used. Version 7.6 used
a newer revision of the format if the network contained non-function nodes with function

48

HUGIN KB files are (as of HUGIN API version 6.2) automatically compressed
using the Zlib library (www.zlib.net). This implies that the developer
(i.e., the user of the HUGIN API) must (on some platforms) explicitly link to
the Zlib library, if the application makes use of HKB files see Section 1.2.
x

h status t h domain save as kb

(h domain t domain, h string t file name, h string t password)
Save domain as a HUGIN KB to a file named file name. If password is not
NULL , then the HKB file will be protected by the given password (i.e., the
file can only be loaded if this password is supplied to the h kb load domain
function).
DBN: domain must not be a (compiled) DBN runtime domain that has been
triangulated using h domain triangulate dbn(65) .

h domain t h kb load domain

(h string t file name, h string t password)
Load a domain from the HUGIN KB file named file name. If the operation is
successful, a reference to the domain is returned. In case of errors (e.g., if
the file was not created by h domain save as kb), NULL is returned.
If the HKB file is password protected, then the password argument must
match the password used to create the HKB file (if not, the load operation
will fail with an invalid password error). If the HKB file is not protected,
the password argument is ignored.
The name of the file from which a domain was loaded or to which it was
saved is stored within the domain object. The file name used in the most recent load or save operation can be retrieved using h domain get file name(201) .

parents; otherwise, it used the same format as version 7.5. Versions 7.7, 7.8, 8.0, and 8.1 use
a new revision of the format if the network contains function nodes; otherwise, they use the
same format as version 7.5. Versions 8.2 and 8.3 use a new revision of the format if the stateindex operator is used within some expression in the network; otherwise, the HKB revision
used is the same as in version 8.1. HUGIN API 8.3 can load HKB files produced by version
5.0 or any later version up to (at least) version 8.3 but future versions of the HUGIN API
might not.

49

50

Chapter 3

Object-Oriented
Belief Networks and LIMIDs
This chapter provides the tools for constructing object-oriented belief network and LIMID models.
An object-oriented model is described by a class. Like a domain, a class
has an associated set of nodes, connected by links. However, a class may
also contain special nodes representing instances of other classes. A class
instance represents a network. This network receives input through input
nodes and provides output through output nodes. Input nodes of a class are
placeholders to be filled-in when the class is instantiated. Output nodes
can be used as parents of other nodes within the class containing the class
instance.
Object-oriented models cannot be used directly for inference: An objectoriented model must be converted to an equivalent flat model (represented as a domain see Chapter 2) before inference can take place.

3.1

Classes and class collections

Classes are represented as C objects of type h class t.

An object-oriented model is comprised of a set of classes, some of which are
instantiated within one or more of the remaining classes. The type h class
collection t is introduced to represent this set of classes. The HUGIN API
requires that there be no references to classes outside of a class collection
(i.e., the class referred to by a class instance must belong to the same class
collection as the class that contains the class instance).
A class collection is edited as a unit: Modifying parts of the interface of a
class will cause all of its instances to be modified in the same way.
51

3.2

Creating classes and class collections

A class always belongs to a (unique) class collection. So, before a class can
be created, a class collection must be created.
h class collection t h new class collection (void)

Create a new empty class collection.

h class t h cc new class (h class collection t cc)

Create a new class. The new class will belong to class collection cc.
h class t h cc get members (h class collection t cc)

Retrieve the list of classes belonging to class collection cc. The list is a NULLterminated list.
h class collection t h class get class collection (h class t class)

Retrieve the class collection to which class class belongs.

3.3

Deleting classes and class collections

h status t h cc delete (h class collection t cc)

Delete class collection cc. This also deletes all classes belonging to cc.
h status t h class delete (h class t class)

Delete class class and remove it from the class collection to which it belongs. If class is instantiated, then this operation will fail. (The h class get
instances(55) function can be used to test whether class is instantiated.)

3.4

Naming classes

In order to generate textual descriptions of classes and class collections in

the form of NET files, it is necessary to name classes.
x

h status t h class set name (h class t class, h string t name)

Set (or change) the name of class to name. The name must be a valid name
(i.e., a valid C identifier) and must be distinct from the names of the other
classes in the class collection to which class belongs.
52

h string t h class get name (h class t class)

Retrieve the name of class. If class is unnamed, a new unique name will
automatically be generated and assigned to class.
h class t h cc get class by name
(h class collection t cc, h string t name)

Retrieve the class with name name in class collection cc. If no such class
exists in cc, NULL is returned.

3.5

Creating basic nodes

Creating basic (i.e., non-instance) nodes in classes is similar to the way

nodes are created in domains (see Section 2.3).
h node t h class new node
(h class t class, h node category t category,
h node kind t kind)

Create a new basic node of the indicated category and kind within class.
The node will have default values assigned to its attributes: The desired
attributes of the new node should be explicitly set using the relevant API
functions.
h class t h node get home class (h node t node)

Retrieve the class to which node belongs. If node is NULL, or node does not
belong to a class (i.e., it belongs to a domain), NULL is returned.
Deletion of basic nodes is done using h node delete(31) .

3.6

Naming nodes

Nodes belonging to classes can be named, just like nodes belonging to domains. The functions to handle names of class nodes are the same as those
used for domain nodes (see Section 2.7) plus the following function.
x

h node t h class get node by name (h class t class, h string t name)

Retrieve the node with name name in class. If no node with that name exists
in class, or if an error occurs (i.e., class or name is NULL), NULL is returned.
53

3.7

The interface of a class

A class has a set of input nodes and a set of output nodes. These nodes
represent the interface of the class and are used to link instances of the class
to other class instances and network fragments.
For the following functions, when a node appears as an argument, it must
belong to a class. If not, a usage error is generated.
x

h status t h node add to inputs (h node t node)

Add node to the set of input nodes associated with the class to which node
belongs. The following restrictions must be satisfied: node must not be an
instance node, node must not be an output node, node must not be an output
clone, and node must not have parents. The last condition is also enforced
by h node add parent(33) it will not add parents to input nodes.
DBN: node must not belong to a class containing temporal clones.

h node t h class get inputs (h class t class)

Retrieve the list of input nodes associated with class.

h status t h node remove from inputs (h node t node)

Remove node from the set of input nodes associated with the class to which
node belongs. (It is checked that node is an input node.) Input bindings
(see Section 3.9) involving node in the instances of the class to which node
belongs are deleted.

h status t h node add to outputs (h node t node)

Add node to the set of output nodes associated with the class to which node
belongs. The following restrictions must be satisfied: node must not be an
instance node, and node must not be an input node. Output clones corresponding to node are created for all instances of the class to which node
belongs.
DBN: node must not belong to a class containing temporal clones.

h node t h class get outputs (h class t class)

Retrieve the list of output nodes associated with class.

h status t h node remove from outputs (h node t node)

Remove node from the set of output nodes associated with the class to which
node belongs. (It is checked that node is an output node.) All output clones
corresponding to node are deleted (if any of these output clones are output
nodes themselves, their clones are deleted too, recursively).
54

This function illustrates that modifying one class may affect many other
classes. This can happen when a class is modified such that its interface,
or some attribute of a node in the interface, is changed. In that case, all
instances of the class are affected. It is most efficient to specify the interface
of a class before creating instances of it.
Deletion of an interface node (using h node delete(31) ) implies invocation of
either h node remove from inputs or h node remove from outputs, depending on whether the node to be deleted is an input or an output node, respectively.

3.8

Creating instances of classes

A class can be instantiated within other classes. Each such instance is represented by a so-called instance node. Instance nodes are of category h
category instance.
x

h node t h class new instance (h class t C1 , h class t C2 )

Create an instance of class C2 . An instance node representing this class
instance is added to class C1 . The return value is this instance node.
DBN: Class C2 must not contain temporal clones.
Output clones (see below) corresponding to the output nodes of class C2 are
also created and added to class C1 .
The classes C1 and C2 must belong to the same class collection. This ensures
that dependencies between distinct class collections cannot be created.
Note that instance nodes define a part-of hierarchy for classes: classes
containing instances of some class C are parents of C. This hierarchy must
form an acyclic directed graph. The h class new instance function checks
this condition. If the condition is violated, or memory is exhausted, the
function returns NULL.
The h node delete(31) function is used to delete instance nodes. Deleting
an instance node will also cause all output clones associated with the class
instance to be deleted (see below).

h class t h node get instance class (h node t instance)

Retrieve the class of which the instance node instance is an instance. (That
is, the class passed as the second argument to h class new instance when
instance was created.)

h node t h class get instances (h class t class)

Retrieve a NULL-terminated list of all instances of class (the list contains an
instance node for each instance of class).
Note that the instance nodes do not belong to class.
55

Output clones
Whenever a class instance is created, instances of all output nodes of the
class are also created. These nodes are called output clones. Since several
instances of some class C can exist in the same class, we need a way to distinguish different copies of some output node Y of class C corresponding to
different instances of C the output clones serve this purpose. For example, when specifying output Y of class instance I as a parent of some node,
the output clone corresponding to the (I, Y) combination must be passed
to h node add parent(33) . Output clones are retrieved using the h node get
output(57) function.
Many API operations are not allowed for output clones. The following restrictions apply:
Output clones can be used as parents, but cannot have parents themselves.
Output clones do not have tables or models.
For discrete output clones, the category and attributes related to states
(i.e., subtype, number of states, state labels, and state values) can be
retrieved, but not set. These attributes are identical to those of the
real output node (known as the master node) and change automatically whenever the corresponding attributes of the master are modified. For example, when the number of states of an output node is
changed, then all tables in which one or more of its clones appear will
automatically be resized as described in Section 2.5.
An output clone cannot be deleted directly. Instead, it is automatically
deleted when its master is deleted or removed from the class interface
(see h node remove from outputs(54) ), or when the class instance to
which it is associated is deleted.
Output clones are created without names, but they can be named just like
other nodes.
An output clone belongs to the same class as the instance node with which
it is associated. Hence, it appears in the node list of that class (and will be
seen when iterating over the nodes of the class).
x

h node t h node get master (h node t node)

Retrieve the output node from which node was cloned. If node is not an
output clone, NULL is returned.
Note that the master itself can be an output clone (since h node add to
outputs(54) permits output clones to be output nodes).
56

h node t h node get instance (h node t node)

Retrieve the instance node with which node is associated. If node is not an
output clone, NULL is returned.
h node t h node get output (h node t instance, h node t output)

Retrieve the output clone that was created from output when instance was
created. (This implies that output is an output node of the class from which
instance was created, and that output is the master of the output clone returned.)
h status t h node substitute class (h node t instance, h class t new)

Change the class instance instance to be an instance of class new. Let old be
the original class of instance. Then the following conditions must hold:
for each input node in old, there must exist an input node in new with
the same name, category, and kind;
for each output node in old, there must exist a compatible output node
in new with the same name.
(Note that this implies that interface nodes must be named.) The notion of
compatibility referred to in the last condition is the same as that used by
h node switch parent(34) and for input bindings (see Section 3.9 below).
The input bindings for instance are updated to refer to input nodes of class
new instead of class old (using match-by-name).
Similarly, the output clones associated with instance are updated to refer to
output nodes of class new instead of class old (again using match-by-name).
This affects only the value returned by h node get master(56) in all other
respects, the output clones are unaffected.
Extra output clones will be created, if class new has more output nodes than
class old.

3.9

Putting the pieces together

In order to make use of class instances, we need to specify inputs to them

and use their outputs. Using their outputs is simply a matter of specifying
the outputs (or, rather, the corresponding output clones) as parents of the
nodes that actually use these outputs. Inputs to class instances are specified
using the following function.
x

h status t h node set input

(h node t instance, h node t input, h node t node)
This establishes an input binding: node is the node to be used as actual input
for the formal input node input of the class of which instance is an instance;
57

instance and node must belong to the same class, and input and node must
be of the same category and kind.
The h node set input function does not prevent the same node from being
bound to two or more input nodes of the same class instance. However, it
is an error if a node ends up being parent of some other node twice in the
runtime domain (Section 3.10). This happens if some node A is bound to
two distinct input nodes, X1 and X2 , of some class instance I, and X1 and X2
have a common child in the class of which I is an instance. This will cause
h class create domain(58) to fail.
Note that for a given input binding to make sense, the formal and actual
input nodes must be compatible. The notion of compatibility used for this
purpose is the same as that used by the h node switch parent(34) and h node
substitute class(57) functions. This means that the nodes must be of the same
category and kind, and (if the nodes are discrete) have the same subtype,
the same number of states, the same list of state labels, and the same list of
state values (depending on the subtype). Only the category/kind restriction
is checked by h node set input. All restrictions (including the category/kind
restriction) are checked by h class create domain(58) .
x

h node t h node get input (h node t instance, h node t input)

For the class instance represented by the instance node instance, retrieve
the actual input node bound to the formal input node input (which must be
an input node of the class of which instance is an instance). If an error is
detected, or no node is bound to the specified input node, NULL is returned.

h status t h node unset input (h node t instance, h node t input)

Delete the input binding (if any) for input in class instance instance (input
must be an input node of the class of which instance is an instance).

3.10

Creating a runtime domain

Before inference can be performed, a class must be expanded to its corresponding flat domain known as the runtime domain.
x

h domain t h class create domain (h class t class)

Create the runtime domain corresponding to class.1 The runtime domain
is not compiled it must be explicitly compiled before it can be used for
inference.
1
For dynamic models, a special runtime domain (known as a DBN runtime domain) is
used. See h class create dbn domain(65) .

58

Creating a runtime domain is a recursive process: First, domains corresponding to the instance nodes within class are constructed (using h class
create domain recursively). These domains are then merged into a common
domain, and copies of all non-instance nodes of class are added to the domain. Finally, the copies of the formal input nodes of the subdomains are
identified with their corresponding actual input nodes, if any.
Note that the runtime domain contains only basic (i.e., non-instance) nodes.
The attributes of the runtime domain are copies of those of class.
Models and tables are copied to the runtime domain. In particular, if tables
are up-to-date with respect to their models in class, then this will also be the
case in the runtime domain. This can save a lot of time (especially if many
copies of a class are made), since it can be very expensive to generate a
table. Generating up-to-date tables is done using h class generate tables(96) .
In order to associate a node of the runtime domain with (the path of) the
node of the object-oriented model from which it was created, a list of nodes
(called the source list) is provided. This node list traces a path from the root
of the object-oriented model to a leaf of the model. Assume the source list
corresponding to a runtime node is hN1 , ..., Nm i. All nodes except the last
must be instance nodes: N1 must be a node within class, and Ni (i > 1) must
be a node within the class of which Ni1 is an instance.
The nodes of the runtime domain are assigned names based on the source
lists: If the name of node Ni is ni , then the name of the runtime node is the
dotted name n1 .n2 . .nm . Because the names of the source nodes are
not allowed to contain dots, this scheme will generate unique (and meaningful) names for all runtime nodes. (As a side-effect of this operation, the
source nodes are also assigned names if they are not already named.)
DBN: In ordinary runtime domains, the nodes have unique source lists. This
is not the case in DBN runtime domains, where the source lists only uniquely
identify nodes in a given time slice. In order to provide unique names for
the nodes, a time slice ID is included in the names. See h class create dbn
domain(65) for more information.
x

h node t h node get source (h node t node)

Return the source list of node. If node doesnt have a source list (i.e., if node
doesnt belong to a runtime domain), or node is NULL (this is an error condition), NULL is returned.
Each node in the source list belongs to some class of the object-oriented
model from which the runtime domain was created.
Note that the contents of the source list will in general be invalidated when
some class of the object-oriented model is modified.
Example 3.1 Consider the object-oriented model shown in Figure 3.1. It has three
basic nodes, A, B, and C, and two instance nodes, I1 and I2 , which are instances of

59

A
I1 :

I2 :

Figure 3.1: An object-oriented belief network model.

the same class. This class has two input nodes, X and Y, and one output node, Z.
Input X of class instance I1 is bound to A. Input Y of class instance I1 and input X
of class instance I2 are both bound to B. Input Y of class instance I2 is unbound.
The runtime domain corresponding to this object-oriented model is shown in Figure 3.2. Note that bound input nodes do not appear in the runtime domain: The
children of a bound input node instead become children of the node to which the
input node is bound. Unbound input nodes, on the other hand, do appear in the
runtime domain.
The node lists returned by h node get source(59) for each node of the runtime domain are as follows: A0 : hAi, B0 : hBi, C0 : hCi, W1 : hI1 , Wi, Z1 : hI1 , Zi, Y2 : hI2 , Yi,
W2 : hI2 , Wi, Z2 : hI2 , Zi.

3.11

Node iterator

In order to iterate over the nodes of a class, the following function is needed.
x

h node t h class get first node (h class t class)

Return a pointer to the first node of class, or NULL if class contains no nodes,
or class is NULL (this is considered an error).
This function should be used in conjunction with the h node get next(43)
function.

60

A0

B0

Y2

W1

W2

Z1

Z2

C0
Figure 3.2: A runtime domain corresponding to the object-oriented model
shown in Figure 3.1.

3.12

User data

Section 2.9 describes functions for associating user-defined data with domains and nodes. Similar functions are also provided for classes.
The first two functions manage generic pointers to data structures that must
be maintained by the user application.
x

h status t h class set user data (h class t class, void data)

Store the pointer data within class.

void h class get user data (h class t class)

Retrieve the value stored within the user data slot of class. If no value has
been stored, the stored value is NULL, or class is NULL (this is an error), NULL
is returned.
The following functions manage key/value-type attributes.

h status t h class set attribute

(h class t class, h string t key, h string t value)
Insert (or update, if key is already defined) the key/value-pair in the attribute
list for class (key must be a valid C language identifier). If value is NULL, the
attribute is removed.
61

h string t h class get attribute (h class t class, h string t key)

Lookup the value associated with key in the attribute list for class. If key is
not present, or if an error occurs, NULL is returned.
This function is needed for iterating over the attributes of a class.

h attribute t h class get first attribute (h class t class)

Retrieve the first attribute object for class. If the attribute list is empty (or
class is NULL), NULL is returned.
The remaining functions needed for iteration over attributes are described
in Section 2.9.

3.13

Saving class collections as HKB files

A class collection can be saved as a HUGIN Knowledge Base (HKB) file. This
is a portable binary file format, which is only intended to be read by the
appropriate HUGIN API functions. There are two types of HKB files: HKB
files containing domains (see Section 2.10) and HKB files containing class
collections.
x

h status t h cc save as kb
(h class collection t cc, h string t file name,
h string t password)
Save cc as a HUGIN KB to a file named file name. If password is not NULL,
then the HKB file will be protected by the given password (i.e., the file can
only be loaded if this password is supplied to the h kb load class collection
function).

h class collection t h kb load class collection

(h string t file name, h string t password)
Load a class collection from the HUGIN KB file named file name. If the operation is successful, a reference to the class collection is returned. In case of
errors (e.g., if the file was not created by h cc save as kb), NULL is returned.
If the HKB file is password protected, then the password argument must
match the password used to create the HKB file (if not, the load operation
will fail with an invalid password error). If the HKB file is not protected,
the password argument is ignored.
The name of the file from which a class collection was loaded or to which it
was saved is stored within the class objects of the class collection. The file
name used in the most recent load or save operation can be retrieved using
h class get file name(201) .
62

Chapter 4

Dynamic Bayesian Networks

Dynamic Bayesian (or Belief) Networks (DBNs) are used to model systems
that evolve over time. The state of the system at a single time instant is
modeled using an ordinary static Bayesian network. By linking together a
number of these networks, a dynamic model is obtained. The part of the
network corresponding to a single time instant is called a time slice [17].

4.1

Temporal clones

The structure of a DBN at a given time instant is modeled using a class. This
class is then instantiated a specified number of times using h class create
dbn domain(65) . The result is called a DBN runtime domain, and it represents the time window of the system. The time window is where evidence is
entered and inference is performed.
When modeling a system that evolves over time, random variables at a given
time instant typically depend on the state of the system at past time instants.
In order to specify such temporal dependencies, temporal clones can be constructed for the regular nodes of the class. A temporal clone of a variable X
represents X at the previous time instant.1 Links can then be added from
temporal clones to the regular nodes.
Hence, the temporal clones represent the interface between two successive
time slices.
x

h node t h node create temporal clone (h node t node)

This function creates a temporal clone of node (which shall hereafter be
referred to as the temporal master).
The following conditions apply:
1
Hugin restricts temporal dependencies to successive time instants. This means that the
models are so-called first order models.

63

 node must not be an instance node, but it must belong to a class.

This class must not have input or output nodes, and it must not be
instantiated in other classes.
node must not be an output clone or a temporal clone, and node must
not already have a temporal clone (i.e., at most one clone can be created for each regular node).
Temporal clones are somewhat similar to output clones in the sense that
they both serve as placeholders for the real thing. The following highlights
the similarities and differences:
The links of the network must respect the natural flow of time. This
means that temporal clones can only have other temporal clones (not
regular nodes) as parents.
Temporal clones can have tables and models. These tables are used to
define the joint distribution over the temporal clones of the first time
slice. (Because links are allowed between temporal clones, any distribution can be specified.) For subsequent time slices, the temporal
clones are replaced by the actual nodes in the previous time slice.
For discrete temporal clones, the category and attributes related to
states cannot be changed directly. Changes must be performed through
their temporal masters: Identical changes are then automatically performed for the clones.
A temporal clone can be deleted. But if the temporal master of the
clone is deleted, then the clone is also deleted.
h node t h node get temporal clone (h node t node)

Retrieve the temporal clone of node. If node does not have a temporal clone,
NULL is returned.
And this is the inverse function:
h node t h node get temporal master (h node t clone)

Retrieve the temporal master of clone. If clone is not a temporal clone, NULL
is returned.

4.2

DBN runtime domains

As for object-oriented models in general, the time slice class must be instantiated (one instance per time slice) to form a so-called DBN runtime domain.
This domain represents the time window, and this is where evidence is entered and inference is performed.
64

h domain t h class create dbn domain

(h class t class, size t number of slices)

This function creates a DBN runtime domain by instantiating class number

of slices times.2 The nodes corresponding to the temporal clones (except
those of the first time slice) are replaced by the correct nodes in the previous
time slices, thereby linking the slices together to form a connected network.3
This construction implies that there will be several runtime nodes with the
same source list (unless there is only one time slice). For each regular node
in class, the DBN runtime domain contains number of slices runtime nodes
with the same source list.
The nodes are named the same way as h class create domain(58) does, except
that the names are prefixed by Thtimei., where htimei ranges from 1 to
number of slices. The names of the temporal clones associated with the first
time slice are the same as the names of their temporal masters, except that
the prefix is T0..

4.3

Inference in DBNs

Before inference can be performed, the DBN runtime domain must be compiled. A DBN runtime domain can be compiled as an ordinary domain. This
produces a compiled domain that only allows exact inference to be performed within the time window itself. The operation of moving the time
window to include future time instants (see below), and the operation of
prediction, can only be performed approximately with this domain. If the
operations must be performed exactly (i.e., without approximations), then
the triangulation must respect extra properties. This is ensured by using the
following function for triangulation.
h status t h domain triangulate dbn
(h domain t domain, h triangulation method t tm)

This function is similar to h domain triangulate(107) except that the interface between successive time slices are made complete. This allows exact
inference in the situations described above to be made.
It is assumed that domain was created by h class create dbn domain(65) , and
that no changes have been made to domain since its creation (in particular, no node creations or deletions, no link changes, and no changes to the
CPTs4 ). The same is assumed for the class from which domain was created
(the class must not be deleted, or modified, until domain is deleted).
2

This is done using h class create domain(58) .

Of course, it is also possible to form a collection of disconnected networks, but that probably wouldnt be of much use.
4
This includes EM learning and adaptation.
3

65

All interface nodes of the network must be discrete chance nodes, and the
network must not contain decision or utility nodes. Also, the network must
not contain a functional trail between any pair of interface nodes.
A (compiled) DBN runtime domain that has been triangulated using h domain triangulate dbn cannot be compressed.5 Also, such domains can only
use sum-propagations for inference. And, in the current implementation,
such domains cannot be saved as HKB files.
If the time window doesnt cover all time instants of interest, it can be moved
using the following function.
x

h status t h domain move dbn window

(h domain t domain, size t )
This operation (which assumes that domain is a compiled DBN runtime domain) moves the time window forward by steps. The number of steps
must be positive and less than or equal to the width of the time window.
Evidence for nodes corresponding to time instants that disappear from the
time window becomes frozen (i.e., it cannot be retracted or altered). But
the evidence is not forgotten it is incorporated in the beliefs of the nodes
in future time slices. This is done exactly if the triangulation of domain was
done using h domain triangulate dbn(65) . Otherwise, approximate inference
is performed.
All interface nodes of the network must be discrete chance nodes, and the
network must not contain decision or utility nodes. Also, the network must
not contain a functional trail between any pair of interface nodes.
The function doesnt establish equilibrium an explicit propagation operation is required. The reason for this choice is to allow evidence to be entered
for the nodes corresponding to new time instants that become visible when
the window is moved before performing a propagation. (An additional propagation would just be a waste of time.)

h count t h domain get dbn window offset (h domain t domain)

Return the total number of time steps that the time window of domain has
been moved.

h status t h domain initialize dbn window (h domain t domain)

Retract evidence for all nodes of domain, move the time window back to its
initial position, and establish the initial state of the inference engine.
5
The reason is that h domain move dbn window uses h domain get marginal(126) (which
doesnt support compressed domains).

66

4.4

Prediction

We can compute beliefs for discrete nodes, means and variances for continuous nodes, and values for real-valued function nodes in time slices that lie
beyond the time window. This operation is called prediction.
h status t h domain compute dbn predictions
(h domain t domain, size t number of time instants)

This operation computes predictions for all nodes at time instants following the time window. Let the time of the last slice in the time window be t.
Then this function computes predictions for all nodes at times t + 1, . . . ,
t + number of time instants. These predictions are referred to using time
indexes ranging from 0 to number of time instants 1.
The number of time instants must be a positive number.
If domain (which must be a compiled DBN runtime domain) is triangulated
using h domain triangulate dbn(65) , the predictions are performed exactly.
Otherwise, approximate inference is performed.
All interface nodes of the network must be discrete chance nodes, and the
network must not contain decision or utility nodes. Also, the network must
not contain a functional trail between any pair of interface nodes.
The junction tree potentials must be up-to-date with respect to the evidence,
the node tables and their models (if any). The equilibrium must be sum,
and the evidence incorporation mode must be normal. This can be ensured
using a propagation operation.
The predictions computed by h domain compute dbn predictions can be accessed by the following functions. In all of these functions, the node argument must be the representative that belongs to the first (complete) time
slice of the time window (i.e., the node that is named with the prefix T1.
by h class create dbn domain(65) ).
The predictions are (only) available as long as domain is compiled.6 Also,
note that the predictions are not automatically updated by propagation operations an explicit call to h domain compute dbn predictions is required.
h number t h node get predicted belief
(h node t node, size t s, size t time)

Retrieve the belief of node for state s at time instant time as computed by the
most recent successful call to h domain compute dbn predictions (provided
that the predictions are still available); node must be a discrete node, and
time must be less than the number of time instants specified in that call.
6

Recall that many HUGIN API functions perform an implicit uncompile operation.

67

h number t h node get predicted mean (h node t node, size t time)

Retrieve the mean value of node at time instant time as computed by the
most recent successful call to h domain compute dbn predictions (provided
that the predictions are still available); node must be a continuous node,
and time must be less than the number of time instants specified in that call.
h number t h node get predicted variance
(h node t node, size t time)

Retrieve the variance of node at time instant time as computed by the most
recent successful call to h domain compute dbn predictions (provided that
the predictions are still available); node must be a continuous node, and
time must be less than the number of time instants specified in that call.
h number t h node get predicted value (h node t node, size t time)

Retrieve the value of node at time instant time as computed by the most
recent successful call to h domain compute dbn predictions (provided that
the predictions are still available); node must be a real-valued function node,
and time must be less than the number of time instants specified in that call.

4.5

The Boyen-Koller approximation algorithm

Boyen and Koller ([4] and [19, 15.3.2]) have described a way to approximate the joint distribution of the interface variables between two successive
time slices such that the accumulated errors remain bounded indefinitely.
The idea is to use a factorization of the distribution so that inference in the
junction tree corresponding to a single time slice has tractable complexity.
During inference, the distribution over the interface variables is approximated according to the desired factorization, and the factors are transferred
to the junction corresponding to the next time slice.
The Hugin inference engine uses this technique to approximate moving of
the time window and prediction when the DBN runtime domain has not
been compiled for exact inference.
The Hugin inference engine uses the links between the temporal clones as
the definition of the factorization to be used for the approximation.

68

Chapter 5

Tables
Tables are used within HUGIN for representing the conditional probability,
policy, and utility potentials of nodes, the probability and utility potentials
on separators and cliques of junction trees, evidence potentials, etc.
The HUGIN API makes (some of) these tables accessible to the programmer
via the opaque pointer type h table t and associated functions.
The HUGIN API currently does not provide means for the programmer to
construct his own table objects, just the functions to manipulate the tables
created by HUGIN.

5.1

What is a table?

A potential is a function from the state space of a set of variables into the set
of real numbers. A table is a computer representation of a potential.
The HUGIN API introduces the opaque pointer type h table t to represent
table objects.
Consider a potential defined over a set of nodes. In general, the state space
of the potential has both a discrete part and a continuous part. Both parts
are indexed by the set I of configurations of states of the discrete nodes.
The discrete data are comprised of numbers x(i) (i I). If the potential is a
probability potential, x(i) is a probability (i.e., a number between 0 and 1,
inclusive). If the potential is a utility potential, x(i) can be any real number.
Probability potentials with continuous nodes represent so-called CG potentials (see [8, 20, 23]). They can either represent conditional or marginal
probability distributions. CG potentials of the conditional type are accessed
using special functions see Section 2.6. For CG potentials of the marginal
type, we have, for each i I, a number x(i) (a probability), a mean value
vector (i), and a (symmetric) covariance matrix (i); (i) and (i) are the
69

mean value vector and the covariance matrix for the conditional distribution
of the continuous nodes given the configuration i of the discrete nodes.
To be able to use a table object effectively, it is necessary to know some facts
about the representation of the table.
The set of configurations of the discrete nodes (i.e., the discrete state space I)
is organized as a multi-dimensional array in row-major format. Each dimension of this array corresponds to a discrete node, and the ordered list of
dimensions defines the format as follows.1
Suppose that the list of discrete nodes is hN1 , . . . , Nn i, and suppose that
node Nk has sk states. A configuration of the states of these nodes is a list
hi1 , . . . , in i, with 0 ik < sk (1 k n).
The set of configurations is mapped into the index set {0, . . . , S 1} where
n
Y

S=

sk

k=1

(This quantity is also known as the size of the table.)

A specific configuration hi1 , . . . , in i is mapped to the index value
n
X

ak i k

k=1

where
ak = sk+1 sn
(Note that this mapping is one-to-one.)
Example 5.1 The mapping from state configurations to table indexes can be expressed using a simple loop. Let node count be the number of discrete nodes in the
given table, let state count[k] be the number of states of the kth node , and let configuration[k] be the state of the kth node in the state configuration. Then the table
index corresponding to the given state configuration can be computed as follows:
size_t k, index = 0;
for (k = 0; k < node_count; k++)
index = index * state_count[k] + configuration[k];
An API function is also provided to perform this computation: h table get index
from configuration(71) .

Many HUGIN API functions use the index of a configuration whenever the
states of a list of discrete nodes are needed. Examples of such functions are:
h node set alpha(41) , h node get alpha(42) , h table get mean(73) , etc.
1
This only applies to uncompressed tables. If a table is compressed, then there is no simple
way to map a configuration to a table index. (Compression is described in Section 7.6.)

70

As a convenience, the HUGIN API provides functions to convert between

table indexes and state configurations.
x

size t h table get index from configuration

(h table t table, const size t configuration)
Compute the table index corresponding to the state configuration specified
in the configuration argument. This must be a list containing state indexes
for the (discrete) nodes of table the kth element of configuration must be
a state index for the kth node in the node list of table. A state index for each
discrete node of table must be specified (so the length of the configuration
list must be at least the number of discrete nodes of table).
If an error occurs, (size t) 1 is returned.

h status t h table get configuration from index

(h table t table, size t configuration, size t index)
Compute the state configuration corresponding to the specified table index.
The configuration is stored in the configuration list (this list must be allocated by the user application). The function is the inverse of the above
function, so the kth element of configuration is the state index corresponding to the kth node in the node list of table. The length of the configuration
list must be at least the number of discrete nodes of table.
Example 5.2 Given three discrete nodes, A with 2 states (a0 and a1 ), B with
3 states (b0 , b1 , and b2 ), and C with 4 states (c0 , c1 , c2 , and c3 ), here is a complete
list of all configurations of hA, B, Ci and their associated indexes:
ha0 , b0 , c0 i, index 0;
ha0 , b0 , c1 i, index 1;
ha0 , b0 , c2 i, index 2;
ha0 , b0 , c3 i, index 3;
ha0 , b1 , c0 i, index 4;
ha0 , b1 , c1 i, index 5;
ha0 , b1 , c2 i, index 6;
ha0 , b1 , c3 i, index 7;
ha0 , b2 , c0 i, index 8;
ha0 , b2 , c1 i, index 9;
ha0 , b2 , c2 i, index 10;
ha0 , b2 , c3 i, index 11;
ha1 , b0 , c0 i, index 12;
ha1 , b0 , c1 i, index 13;
ha1 , b0 , c2 i, index 14;
ha1 , b0 , c3 i, index 15;
ha1 , b1 , c0 i, index 16;

71

ha1 , b1 , c1 i, index 17;

ha1 , b1 , c2 i, index 18;
ha1 , b1 , c3 i, index 19;
ha1 , b2 , c0 i, index 20;
ha1 , b2 , c1 i, index 21;
ha1 , b2 , c2 i, index 22;
ha1 , b2 , c3 i, index 23.

5.2
x

The nodes and the contents of a table

h node t h table get nodes (h table t table)

Retrieve the NULL-terminated list of nodes associated with table. If an error

occurs, NULL is returned.
The first part of this list is comprised of the discrete nodes of the potential
represented by table. The order of these nodes determines the layout of
the discrete state configurations as described in the previous section. The
second part of the list is comprised of the continuous nodes of the potential
represented by table.
The pointer returned by h table get nodes is a pointer to the list stored in the
table structure. Do not modify or deallocate it.
x

h number t h table get data (h table t table)

Retrieve a pointer to the array of table holding the actual discrete data (denoted by x(i) in Section 5.1). This array is a one-dimensional (row-major)
representation of the multi-dimensional array.
Since the pointer returned is a pointer to the actual array stored within the
table structure, it is possible to modify the contents of the table through this
pointer. But, of course, the pointer must not be deallocated.
Note that pointers to nodes and data arrays within tables may be invalidated
by other API functions. For example, suppose the application holds a pointer
to the data array of the conditional probability table (CPT) of some node X.
Changing the set of parents of X (or the number of states of X or one of its
parents) causes the CPT to be resized. Most likely, this will cause the data
array of the CPT to be moved to a different memory location, and the old
pointer is no longer valid.
For tables with continuous nodes, h table get data is used to access the x(i)
component. To access the (i) and (i) components, the following functions
72

must be used (we assume that table is a table returned by h domain get
marginal(126) or h node get distribution(126) ):
h double t h table get mean (h table t table, size t i, h node t node)

Return the mean value of the conditional distribution of the continuous node
node given the discrete state configuration i.
h double t h table get covariance
(h table t table, size t i, h node t node1 , h node t node2 )

Return the covariance of the conditional distribution of the continuous nodes

node1 and node2 given the discrete state configuration i.
h double t h table get variance
(h table t table, size t i, h node t node)

Return the variance of the conditional distribution of the continuous node

node given the discrete state configuration i.

5.3

Deleting tables

The HUGIN API also provides a function to release the storage resources
used by a table. The h table delete function can be used to deallocate tables returned by h domain get marginal(126) , h node get distribution(126) , h
node get experience table(160) , and h node get fading table(161) . All other deletion requests are ignored (e.g., a table returned by h node get table(39)
cannot be deleted).
h status t h table delete (h table t table)

Release the memory resources used by table.

5.4

The size of a table

The number of state configurations of discrete nodes represented in a table

is referred to as the size of the table. If the table has n discrete nodes, and
the kth node has sk states, then the size of the table is
n
Y
sk
k=1

This assumes that all state configurations are represented in the table. If
some state configurations have been removed (by a process known as compression see Section 7.6), the size will be smaller.
x

size t h table get size (h table t table)

Return the size of table. If an error occurs, (size t) 1 is returned.
73

Associated with each state configuration of discrete nodes in the table is a

real number of type h number t. In a single-precision version of the HUGIN
API, h number t is a 4-byte quantity, but in a double-precision version, it is
an 8-byte quantity.
If the table has continuous nodes, then there will be additional data stored
in the table (such as mean and covariance values). If the table has m continuous nodes, the number of additional data elements is
m(m + 3) Y
sk
2
n

k=1

We call this quantity the CG size of the table. (If the table is compressed,
then this quantity is reduced proportional to the number of discrete configurations removed from the table.)
size t h table get cg size (h table t table)

Return the CG size of table. If an error occurs, (size t) 1 is returned.

Each CG data element occupies 8 bytes.

5.5

Rearranging the contents of a table

Sometimes it is convenient to enforce a specific layout of the contents of a

table. This can be done by permuting the node list of the table.
x

h status t h table reorder nodes (h table t table, h node t order)

Reorder the node list of table to be order (the contents of the data arrays
are reorganized according to the new node order); order must be a NULLterminated list containing a permutation of the node list of table. If table is
the node table of some (chance or decision) node, then that node must have
the same position in order as in the node list of table. Also, if the node list of
table contains both discrete and continuous nodes, the discrete nodes must
precede the continuous nodes in order.
If a chance node has experience or fading tables, then the order of the (discrete) parents in these tables must be the same as in the conditional probability table. This constraint is enforced by h table reorder nodes. So, reordering one of these tables also reorders the other two tables (if they exist).
In the current implementation of the HUGIN API, reordering the nodes of
a node table causes the affected domain to be uncompiled. (Except, if the
node list of table is equal to order, then nothing is done.)
Example 5.3 The following code creates four chance nodes, two discrete (a and b)
and two continuous (x and y); a, b, and x are made parents of y. Then the number
of states of a and b and the conditional distributions of the nodes must be specified
(this is not shown).

74

h_domain_t d = h_new_domain ();

h_node_t a = h_domain_new_node (d, h_category_chance,
h_kind_discrete);
h_node_t b = h_domain_new_node (d, h_category_chance,
h_kind_discrete);
h_node_t x = h_domain_new_node (d, h_category_chance,
h_kind_continuous);
h_node_t y = h_domain_new_node (d, h_category_chance,
h_kind_continuous);
h_node_add_parent (y, a);
h_node_add_parent (y, b);
h_node_add_parent (y, x);
...

/* set number of states,

specify conditional distributions, etc. */

Now suppose we want to ensure that a appears before b in the node list of the
conditional probability table of y. We make a list containing the desired order of y
and its parents, and then we call h table reorder nodes.
h_node_t list[5];
h_table_t t = h_node_get_table (y);
list[0] = a; list[1] = b;
list[2] = y; list[3] = x;
list[4] = NULL;
h_table_reorder_nodes (t, list);
Note that since y (the child node of the table) is continuous, it must be the first
node among the continuous nodes in the node list of the table. Had y been discrete,
it should have been the last node in the node list of the table (in this case, all nodes
would be discrete).

75

76

Chapter 6

Generating Tables
This chapter describes how to specify a compact description of a node table.
From this description, the contents of the table is generated.
Such a table description is called a model. A model consists of a list of discrete nodes and a set of expressions (one expression for each configuration
of states of the nodes). The expressions are built using standard statistical distributions (such as Normal, Binomial, Beta, Gamma, etc.), arithmetic
operators (such as addition, subtraction, etc.), standard functions (such as
logarithms, exponential, trigonometric, and hyperbolic functions), logical
operators (conjunction, disjunction, and conditional), and relations (such
as less-than or equals).
Models are also used to specify the functions associated with real-valued
function nodes. In this case, no tables are generated.

6.1

Subtyping of discrete nodes

In order to provide a rich language for specifying models, we introduce a

classification of the discrete nodes into four groups:
Labeled nodes. These are discrete nodes that have a label associated
with each state (and nothing else). Labels can be used in equality
comparisons and to express deterministic relationships.
Boolean nodes. Such nodes represent the truth values, false and true
(in that order).
Numbered nodes. The states of such nodes represent numbers (not
necessarily integers).
Interval nodes. The states of such nodes represent (disjoint) intervals
on the real line.
The last two groups are collectively referred to as numeric nodes.
77

Recall that discrete nodes have a finite number of states. This implies that
numbered nodes can only represent a finite subset of, e.g., the nonnegative
integers (so special conventions are needed for discrete infinite distributions
such as the Poisson see Section 6.7.2).
The above classification of discrete nodes is represented by the enumeration
type h node subtype t. The constants of this enumeration type are: h
subtype label, h subtype boolean, h subtype number, and h subtype interval.
In addition, the constant h subtype error is defined for denoting errors.
h status t h node set subtype
(h node t node, h node subtype t subtype)

Set the subtype of node (which must be a discrete node) to subtype.

OOBN/DBN: node must not be an output clone or a temporal clone.
If subtype is h subtype boolean then node must have exactly two states. Moreover, when a node has subtype boolean, h node set number of states(38)
cannot change the state count of the node.
The default subtype (i.e., if it is not set explicitly using the above function)
of a node is h subtype label.
The state labels and the state values (see h node set state label(88) and h
node set state value(89) ) are not affected by this function.
h node subtype t h node get subtype (h node t node)

Return the subtype of node (which must be a discrete node). If node is NULL
or not a discrete node, h subtype error is returned.

6.2

Expressions

Expressions are classified (typed) by what they denote. There are four different types: labeled, boolean, numeric, and distribution.1
An opaque pointer type h expression t is defined to represent the expressions that constitute a model. Expressions can represent constants, variables, and composite expressions (i.e., expressions comprised of an operator
applied to a list of arguments). The HUGIN API defines the following set of
functions to construct expressions.
All these functions return NULL on error (e.g., out-of-memory).
x

h expression t h node make expression (h node t node)

This function constructs an expression that represents a variable. If node is a
CG, a utility, or a real-valued function node, then the constructed expression
1
In a conditional expression, it is possible to combine expressions of different types
resulting in a union type. See h operator if (82) .

78

is of numeric type. If node is a discrete node, then the type of the expression
is either labeled, boolean, or numeric, depending on the subtype of node.
x

h expression t h label make expression (h string t label)

Construct an expression that represents a label constant. A copy is made of
label.

h expression t h boolean make expression (h boolean t b)

Construct an expression that represents a boolean constant: true if b is 1,
and false if b is 0.

h expression t h number make expression (h double t number)

Construct an expression representing the numeric constant number.
The expressions constructed using one of the above four functions are called
simple expressions, whereas the following function constructs composite expressions.

h expression t h make composite expression

(h operator t operator, h expression t arguments)
This function constructs a composite expression representing operator applied to arguments, which must be a NULL-terminated list of expressions.
The function allocates an internal array to hold the expressions, but it does
not copy the expressions themselves.
The h operator t type referred to above is an enumeration type listing all
possible operators, including statistical distributions.
The complete list is as follows:
h operator add, h operator subtract, h operator multiply,
h operator divide, h operator power
These are binary operators that can be applied to numeric expressions.
h operator negate
A unary operator for negating a numeric expression.
h operator equals, h operator less than, h operator greater than,
h operator not equals, h operator less than or equals,
h operator greater than or equals
These are binary comparison operators for comparing labels, numbers, and boolean values (both operands must be of the same type).
Only the equality operators (i.e., h operator equals and h operator
not equals) can be applied to labels and boolean values.
79

 h operator state index

This operator denotes the index (a nonnegative integer) of the state
of a discrete node (that is, the argument to this operator must be an
expression constructed using h node make expression(78) applied to a
discrete node). This is a numeric expression. In the syntax of expressions, this operator is represented by the # symbol.
All of the above operators are represented by special symbols in the syntax
of expressions, while the following operators are named operators i.e.,
they are written as a function application with a name followed by a list
of arguments.
h operator Normal, h operator LogNormal, h operator Beta,
h operator Gamma, h operator Exponential, h operator Weibull,
h operator Uniform, h operator Triangular, h operator PERT
Continuous statistical distributions see Section 6.7.1.
h operator Binomial, h operator Poisson, h operator NegativeBinomial,
h operator Geometric, h operator Distribution, h operator NoisyOR
Discrete statistical distributions see Section 6.7.2.
h operator truncate
This operator can be applied to a continuous statistical distribution in
order to form a truncated distribution. The operator takes either two
or three arguments. When three arguments are specified, the first and
third arguments must be numeric expressions denoting, respectively,
the left and right truncation points, while the second argument must
denote the distribution to be truncated.
Either the first or the third argument can be omitted. Omitting the
first argument results in a right-truncated distribution, and omitting
the third argument results in a left-truncated distribution.
Example 6.1 Using the syntax described in Section 6.3, a truncated normal
distribution can be expressed as follows:
truncate (-4, Normal (0, 1), 4)
This distribution is truncated at the left at 4 and at the right at 4. A lefttruncated (at 4) normal distribution is obtained by omitting the last argument:
truncate (-4, Normal (0, 1))
Similarly, a right-truncated normal distribution can be obtained by omitting
the first argument.

80

 h operator aggregate
This operator denotes the distribution of the sum of a random number
of independent identically distributed random variables.
The number of random variables (the number of events) is modeled by
a discrete distribution, called the frequency distribution. The severity
of each event is modeled by a continuous distribution. The frequency
and severity distributions are assumed to be independent.
This is known as an aggregate distribution.
The operator takes two arguments: The first argument (representing
the frequency distribution) must be a numbered node, and the state
values for that node must form the sequence 0, 1, . . . , m for some m.
The second argument (representing the severity distribution) must be
an interval node (usually ranging from zero to infinity).2
An aggregate distribution can only be specified for a discrete function node (of interval subtype).2 The reason is that the frequency and
severity distributions must be available before the aggregate distribution can be computed. The intervals of the function node must cover
all values of (the domain of) the aggregate distribution. This usually
means all values from zero to infinity.
Example 6.2 An aggregate distribution can be used to model the total claim
amount for an insurance portfolio. The number of claims is modeled by the
frequency distribution (this might be a Poisson distribution), and the size (or
cost) of each claim is modeled by the severity distribution.

h operator probability
This operator takes a single argument a boolean expression. This argument must contain exactly one discrete node (and no other nodes).
Moreover, the operator cannot be nested.
The expression is evaluated by instantiating the node in the argument
to all possible states. The result is the sum of the beliefs of the instantiations for which the boolean argument evaluates to true. (This means
that inference must be performed before the expression can be evaluated.)
The operator can only be used in models for function nodes.
Example 6.3 If X is a labeled discrete node with states low, medium, and
high, then the following expressions are all valid in models for function nodes
having X as parent.
2

Aggregate distributions are currently not supported for nodes with zero-width intervals.

81

probability (X == "low")
probability (X != "low")
probability (or (X == "low", X == "medium"))
Such expressions can be used to transfer probabilities from one subnetwork
to another if they are connected by a functional trail.

h operator min, h operator max

Compute the minimum or maximum of a list of numbers (the list must
be non-empty).
h operator log, h operator log2, h operator log10, h operator exp,
h operator sin, h operator cos, h operator tan, h operator sinh,
h operator cosh, h operator tanh, h operator sqrt, h operator abs
Standard mathematical functions to compute the natural (i.e., base e)
logarithm, base 2 and base 10 logarithms, exponential, trigonometric
functions, hyperbolic functions, square root, and absolute value of a
number.
h operator floor, h operator ceil
The floor and ceiling functions round real numbers to integers.
floor(x) (usually denoted bxc) is defined as the greatest integer less
than or equal to x.
ceil(x) (usually denoted dxe) is defined as the least integer greater than
or equal to x.
h operator mod
The modulo function gives the remainder of a division. It is defined
as follows:
mod(x, y) x ybx/yc,
y 6= 0.
Note that x and y can be arbitrary real numbers (except that y must
be nonzero).
h operator if
Conditional expression (with three arguments): The first argument
must denote a boolean value, and the second and third arguments
must denote values of compatible types. The simple case is for these
types to be identical: Then the result of the conditional expression is
of that common type.
Else, the type of the conditional expression is a union type. Let T be
any scalar type (i.e., either labeled, boolean, or numeric). The result
of the conditional expression is of type T -or-distribution if the types of
the last two arguments are any two (distinct) members of the list: T ,
distribution, and T -or-distribution.
82

The union-type feature can be used to define relationships that are

sometimes probabilistic and sometimes deterministic.
Example 6.4 Let A be an interval node, and let B be a boolean node and
a parent of A. Assume that the following expression is used to define the
conditional probability table of A.
if (B, Normal (0,1), 0)
If B is true, then A has a normal distribution. If B is false, then A is instantiated to the interval containing 0.

h operator and, h operator or, h operator not

Standard logical operators: not takes exactly one boolean argument,
while and and or take a list (possibly empty) of boolean arguments.
Evaluation of an and composite expression is done sequentially, and
evaluation terminates when an argument that evaluates to false is
found. Likewise for an or expression (except that the evaluation terminates when an argument evaluating to true is found).
In addition, a number of operators are introduced to denote simple expressions and errors (for use by h expression get operator(83) ):
h operator label for expressions constructed using h label make expression(79) ;
h operator number for expressions constructed using h number make
expression(79) ;
h operator boolean for expressions constructed using h boolean make
expression(79) ;
h operator node for expressions constructed using h node make expression(78) ;
h operator error for illegal arguments to h expression get operator(83) .
Analogous to the constructor functions, the HUGIN API provides functions
to test how a particular expression was constructed and functions to access
the parts of an expression.
x

h boolean t h expression is composite (h expression t e)

Test whether the expression e was constructed using h make composite expression(79) .

h operator t h expression get operator (h expression t e)

Return the operator that was used when the expression e was constructed
using h make composite expression(79) , or, if e is a simple expression, one of
the special operators (see the list above).
83

h expression t h expression get operands (h expression t e)

Return the operand list that was used when the expression e was constructed
using h make composite expression(79) . Note that the returned list is the real
list stored inside e, so dont try to deallocate it after use.

h node t h expression get node (h expression t e)

Return the node that was used when the expression e was constructed using
h node make expression(78) .

h double t h expression get number (h expression t e)

Return the number that was used when the expression e was constructed
using h number make expression(79) . If an error occurs (e.g., e was not constructed using h number make expression), a negative number is returned.
However, since negative numbers are perfectly valid in this context, errors
must be checked for using h error code(18) and friends.

h string t h expression get label (h expression t e)

Return the label that was used when the expression e was constructed using
h label make expression(79) . Note that the real label inside e is returned, so
dont try to deallocate it after use.

h boolean t h expression get boolean (h expression t e)

Return the boolean value that was used when the expression e was constructed using h boolean make expression(79) .

h status t h expression delete (h expression t e)

Deallocate the expression e, including subexpressions (in case of composite
expressions).
This function is called automatically in a number of cases: When a new expression is stored in a model (see h model set expression(87) ), when parents
are removed, and when the number of states of a (discrete) node is changed.
Note: The HUGIN API keeps track of the expressions stored in models. This
means that if you delete an expression with a subexpression that is shared
with some expression within some model, then that particular subexpression
will not be deleted.
However, if you build two expressions with a shared subexpression (and
that subexpression is not also part of some expression owned by HUGIN),
then the shared subexpression will not be protected against deletion if you
delete one of the expressions. For such uses, the following function can be
used to explicitly create a copy of an expression.

h expression t h expression clone (h expression t e)

Create a copy of e.
84

6.3

Syntax of expressions

In many situations, it is convenient to have a concrete syntax for presenting

expressions (e.g., in the HUGIN GUI application). The syntax is also used in
specifications written in the NET language (see Chapter 13).
hExpressioni

hSimple expressioni hComparisoni hSimple expressioni

| hSimple expressioni

hSimple expressioni
hSimple expressioni hPlus or minusi hTermi
| hPlus or minusi hTermi
| hTermi
hTermi
hTermi hTimes or dividei hExp factori
| hExp factori
hExp factori

hFactori hExp factori

| hFactori

hFactori

hUnsigned numberi
| hNode namei
| hStringi
| false
| true
| # hNode namei
| ( hExpressioni )
| hOperator namei ( hExpression sequencei )

hExpression sequencei
hEmptyi
| hExpressioni [ , hExpressioni ]*
hComparisoni3

== | = | != | <> | < | <= | > | >=

hPlus or minusi

+ | -

hTimes or dividei * | /
hOperator namei Normal | LogNormal | Beta | Gamma
| Exponential | Weibull
| Uniform | Triangular | PERT
| Binomial | Poisson | NegativeBinomial
| Geometric | Distribution | NoisyOR
| truncate | aggregate | probability
| min | max | log | log2 | log10 | exp
| sin | cos | tan | sinh | cosh | tanh
| sqrt | abs | floor | ceil | mod
| if | and | or | not
85

The operator names refer to the operators of the h operator t(79) type: prefix the operator name with h operator to get the corresponding constant of
the h operator t type.
h expression t h string parse expression
(h string t s, h model t model,
void (error handler) (h location t, h string t, void ),
void data)

Parse the expression in string s and construct an equivalent h expression t

structure. Node names appearing in the expression must correspond to parents of the owner of model. If an error is detected, the error handler function is called with the location (the character index) within s of the error, a
string that describes the error, and data. The storage used to hold the error
message string is reclaimed by h string parse expression, when error handler
returns (so if the error message will be needed later, a copy must be made).
The user-specified data allows the error handler to access non-local data
(and hence preserve state between calls) without having to use global variables.
The h location t type is an unsigned integer type (such as unsigned long).
If no error reports are desired (in this case, only the error indicator returned
by h error code(18) will be available), then the error handler argument may
be NULL.
Note: The error handler function may also be called in non-fatal situations
(e.g., warnings).
h string t h expression to string (h expression t e)

Allocate a string and write into this string a representation of the expression e using the above described syntax.
Note that it is the responsibility of the user of the HUGIN API to deallocate
the returned string when it is no longer needed.

6.4

Creating and maintaining models

The HUGIN API introduces the opaque pointer type h model t to represent
models. Models must be explicitly created before they can be used.
h model t h node new model
(h node t node, h node t model nodes)

Create and return a model for node (which must be a discrete, a utility, or a
function node) using model nodes (a NULL-terminated list of discrete nodes,
3

Note that both C and Pascal notations for equality/inequality operators are accepted.

86

comprising a subset of the parents of node) to define the configurations of

the model. If node already has a model, it will be deleted before the new
model is installed.
OOBN: node must not be an output clone.
If a node has a model (and a node table), the contents of the table will be
generated from the model. This happens automatically as part of compilation, propagation, and reset-inference-engine operations, but it can also be
explicitly done by the user (see h node generate table(95) ). It is possible to
modify the contents generated from the model, but note that the inference
engine will regenerate the table when certain parameters are changed (see
Section 6.8 for precise details).
x

h model t h node get model (h node t node)

Return the model associated with node. If no model exists, NULL is returned.

h status t h model delete (h model t model)

Delete model (i.e., deallocate the storage used by model). If model was a
model for a node table, the table will again be the definitive source (i.e., the
contents of the table will no longer be derived from a model).

h node t h model get nodes (h model t model)

Return the list of nodes of model.

size t h model get size (h model t model)

Return the number of configurations of the nodes of model. If an error
occurs, (size t) 1 (i.e., the number 1 cast to the type size t) is returned.

h status t h model set expression

(h model t model, size t index, h expression t e)
Store the expression e at position index in model. It is an error if model is
NULL , or index is out of range. If there is already an expression stored at
the indicated position, then that expression is deleted using h expression
delete(84) .
Now, let node be the owner of model (i.e., model is associated with node).
If node is a utility or a real-valued function node, then the type of e must be
numeric.
Otherwise, node is a discrete node:
If e evaluates to a distribution (a standard statistical distribution, a
truncated, or an aggregate distribution), then the distribution must be
appropriate for the subtype of node see Section 6.7.
87

 Otherwise, the value of node is a deterministic function of the parents

(for the configurations determined by index). In this case, the type of e
must match the subtype of node.
In all cases, the subexpressions of e must not use node as a variable only
parents can be used as variables.
h expression t h model get expression
(h model t model, size t index)

Return the expression stored at position index within model. If model is NULL
or no expression has been stored at the indicated position, NULL is returned
(the first case is an error).

6.5

State labels

Labels assigned to states of discrete nodes serve two purposes: (1) to identify states of labeled nodes in the table generator, and (2) to identify states in
the HUGIN GUI application (for example, when beliefs or expected utilities
are displayed).
x

h status t h node set state label

(h node t node, size t s, h string t label)
Create a copy of label and assign it as the label of state s of node (which must
be a discrete node). The label can be any string (that is, it is not restricted
in the way that, e.g., node names are).
When node is used as a labeled node with the table generator facility, the
states must be assigned unique labels.
OOBN/DBN: node must not be an output clone or a temporal clone.

h string t h node get state label (h node t node, size t s)

Return the label of state s of node. If no label has been assigned to the state,
a default label is returned. The default label is the empty string, unless node
is a boolean node in which case the default labels are "false" (for state 0)
and "true" (for state 1).
If an error occurs (i.e., node is not a discrete node, or s is an invalid state),
NULL is returned.
Note that the string returned by h node get state label is not a copy. Thus,
the application should not attempt to free it after use.
The following function provides the inverse functionality.

h index t h node get state index from label

(h node t node, h string t label)
Return the index of the state of node matching the specified label. If node is
88

not a discrete node, or label is NULL (these conditions are treated as errors),
or no (unique) state matching the specified label exists, 1 is returned.
The following function can be used to reorder the states of a labeled node.
h status t h node reorder states (h node t node, h string t order)

Reorder the states of node according to the specified order, where node is a
labeled discrete node, and order is a NULL-terminated list of strings containing a permutation of the state labels of node. The states of node must be
uniquely labeled.
OOBN/DBN: node must not be an output clone or a temporal clone.
In addition to reordering the state labels of node, the data in all tables containing node are reorganized according to the new ordering. The affected
tables are the same as those resized by h node set number of states(38) .
If node is a model node in the model of a child, then the expressions in
that model are reorganized according to the new ordering.
OOBN: If node is an output node of a class, then similar updates are performed for all output clones of node. The process is repeated (recursively) if
some output clones are also output nodes.
DBN: If node has a temporal clone, then similar updates are performed for
the temporal clone.
The elements of the vector of findings and the case data4 associated with
node are updated to match the new state ordering.
Unless order is identical to the current ordering of the states of node, and if
node belongs to a domain, then that domain will be uncompiled.
Notice that states can only be reordered for labeled nodes (not numeric or
boolean nodes).

6.6

State values

Similar to the above functions for dealing with state labels, we need functions for associating states with points or intervals on the real line. We
introduce a common set of functions for handling both of these purposes.
h status t h node set state value
(h node t node, size t s, h double t value)

Associate value with state s of node (which must be a numeric node).

OOBN/DBN: node must not be an output clone or a temporal clone.
4

If node has more than 32767 states, then state indexes 32767 in the case data that
should be mapped (by order) to indexes > 32767 are instead set to missing. The reason is
that state indexes in case data are stored as signed 2-byte quantities see Section 12.1.

89

When node is used with the table generator facility, the state values must
form an increasing sequence.
For numbered nodes, value indicates the specific number to be associated
with the specified state.
For interval nodes, the values specified for state i and state i + 1 are the left
and right endpoints of the interval denoted by state i (the dividing point
between two adjacent intervals is taken to belong to the interval to the right
of the point, except when the first interval has zero width see below). To
indicate the right endpoint of the rightmost interval, specify s equal to the
number of states of node.
To specify (semi-)infinite intervals, the constant h infinity is defined. The
negative of this constant may be specified for the left endpoint of the first
interval, and the positive of this constant may be specified for the right
endpoint of the last interval.
As a special case, the left and right endpoints can be equal. Such so-called
zero-width intervals can be used to express distributions that are combinations of discrete and continuous distributions. In this case, both the left and
right endpoints belong to the interval (so the next interval, if any, is open
at the left end).
h double t h node get state value (h node t node, size t s)

Return the value associated with state s for the numeric node node.
The following function provides the inverse functionality.
h index t h node get state index from value
(h node t node, h double t value)

Return the index of the state of node matching the specified value:
If node is an interval node, the state index of the interval containing
value is returned. If an error is detected (that is, if the state values
of node do not form an increasing sequence), or no interval contains
value, 1 is returned.
If node is a numbered node, the index of the state having value as the
associated state value is returned. If no (unique) state is found, 1 is
returned.
If node is not a numeric node (this is an error condition), 1 is returned.

6.7

Statistical distributions

This section defines the distributions that can be specified using the model
feature of HUGIN.
90

6.7.1

Continuous distributions

Continuous distributions are relevant for interval nodes only.

Normal A random variable X has a normal (or Gaussian) distribution with
mean and variance 2 if its probability density function is of the form
pX (x) =

1
22

e 2 (x) /

<x<

2 > 0

This distribution is denoted Normal(, 2 ).5

Log-Normal A random variable X has a log-normal distribution if log X has
a normal distribution with mean and variance 2 :
1
1
2 2
pX (x) =
e 2 (log x) /
x 22

2 > 0

x>0

This distribution is denoted LogNormal(, 2 ).5

An optional location parameter c can be specified if X c has a lognormal distribution: LogNormal(, 2 , c).
Gamma A random variable X has a gamma distribution if its probability
density function is of the form
pX (x) =

(x/b)a1 ex/b
b (a)

a>0

b>0

x0

a is called the shape parameter, and b is called the scale parameter.

This distribution is denoted Gamma(a, b).
An optional location parameter c can be specified if X c has a gamma
distribution: Gamma(a, b, c).
Beta A random variable X has a beta distribution if its probability density
function is of the form
pX (x) =

1
(x a)1 (b x)1
B(, )
(b a)+1

>0

>0

axb

This distribution is denoted Beta(, , a, b). The standard form of the

beta distribution is obtained by letting a = 0 and b = 1. This variant
is denoted by Beta(, ) and has the density
pX (x) =

1
x1 (1 x)1
B(, )

>0

>0

0x1

The normal and log-normal distributions are often specified using the standard deviation instead of the variance 2 . To be consistent with the convention used for CG potentials,
we have chosen to use the variance.

91

Weibull A random variable X has a Weibull distribution if its probability

density function is of the form
a
pX (x) =
b

 a1
x

e(x/b)

a>0

b>0

x0

a is called the shape parameter, and b is called the scale parameter.

This distribution is denoted Weibull(a, b).
An optional location parameter c can be specified if Xc has a Weibull
distribution: Weibull(a, b, c).
Exponential A random variable X has an exponential distribution if its
probability density function is of the form
pX (x) = ex/b/b

b>0

x0

This distribution is denoted Exponential(b).

An optional location parameter c can be specified if X c has an exponential distribution: Exponential(b, c).
Note: This is a special case of the gamma and Weibull distributions
(obtained by letting the shape parameter a = 1).
Uniform A random variable X has a uniform distribution if its probability
density function is of the form
pX (x) =

1
ba

axb

a<b

This distribution is denoted Uniform(a, b).

Triangular A random variable X has a triangular distribution if its probability density function is of the form

x a axm

a<m
ma
2
pX (x) =

amb
a<b
ba
b x mxb

m<b
bm
This distribution is denoted Triangular(a, m, b).
PERT A PERT distribution is a beta distribution specified using the parameters: a (min), m (mode), b (max) (a < m < b), and an optional shape
parameter ( > 0). The mean of the beta distribution is assumed to
be (a + m + b)/( + 2), and from this assumption, formulas for computing the and parameters of the beta distribution can be derived:
=1+

ma
ba

and
92

=1+

bm
ba

This beta distribution is denoted PERT(a, m, b, ). The 3-parameter

variant PERT(a, m, b) is obtained by letting the shape parameter = 4.
This method of specifying a beta distribution is described in [37].
When a continuous distribution is specified for an interval node, the intervals must include the domain specified in the definition of the distribution.
The density is defined to be zero outside the intended domain of the distribution (e.g., the density for a gamma distributed random variable X is zero
for negative x).
Alternatively, the continuous distribution can be truncated to fit the intervals
of the interval node see h operator truncate(80) .
Finally, it is recommended not to make the intervals larger than necessary.
For example, if a node A has a beta distribution, and A has a child B with
a binomial distribution where A is the probability parameter, then the intervals for A should cover the interval [0, 1] and nothing more. Otherwise,
there would be problems when computing the probabilities for B (because
the probability parameter would be out-of-range).

6.7.2

Discrete distributions

The following discrete distributions apply to numbered and interval nodes.

Binomial A random variable X has a binomial distribution with parameters n (a nonnegative integer) and p (a probability) if
!

P(X = k) =

n k
p (1 p)nk
k

k = 0, 1, . . . , n

This distribution is denoted Binomial(n, p).

Poisson A random variable X has a Poisson distribution with parameter
(a positive real number) if
P(X = k) =

e k
k!

>0

k = 0, 1, 2, . . .

This distribution is denoted Poisson().

Negative Binomial A random variable X has a negative binomial distribution with parameters r (a positive real number) and p (a probability)
if
!
k+r1 r
P(X = k) =
p (1 p)k
k = 0, 1, 2, . . .
k
This distribution is denoted NegativeBinomial(r, p).
93

Geometric A random variable X that counts the number of failures in a

sequence of Bernoulli trials before the first success has a geometric
distribution. Let p denote the probability of success, and let q = 1 p.
Then
P(X = k) = pqk
k = 0, 1, 2, . . .
This distribution is denoted Geometric(p).
Note: The geometric distribution is a special case of the negative binomial distribution (corresponding to r = 1).
When one of the above distributions is specified for a numbered node, the
state values for that node must form the sequence 0, 1, . . . , m for some m.
If the distribution is a binomial distribution, all possible outcomes must be
included (i.e., m n). The other distributions have an infinite number of
possible outcomes, so by convention the probability P(X m) is associated
with the last state.
If any of the above distributions is specified for an interval node, the intervals must include all possible outcomes. Recall that the intervals of an
interval node are taken to be of the form [a, b) (except for the rightmost interval and for infinite endpoints), so, for example, the interval [0, 2) contains
the integers 0 and 1.
The Distribution operator specifies a user-defined distribution. It applies
to all discrete nodes.
Distribution If a discrete node A has n states, then Distribution(e1 , . . . , en )
means that expression ei will be evaluated and the result assigned as
the probability of the ith state (the probabilities need not be normalized). The expressions must be of numeric type and must evaluate to
nonnegative values (and at least one of them must be positive).
The Noisy-OR distribution applies to boolean nodes.
Noisy-OR Let b1 , . . . , bn (n 1) be boolean values, and let q1 , . . . , qn
(0 qi 1) be probability values. A random (boolean) variable X has
a Noisy-OR distribution if
P(X = false) =

qi

i:bi =true

This distribution is denoted NoisyOR(b1 , q1 , . . . , bn , qn ).

The Noisy-OR distribution can be used to model an event that may be
caused by any member of a set of conditions, and the likelihood of
causing the event increases if more conditions are satisfied.
The assumptions underlying the Noisy-OR distribution are:
94

 If all the causing conditions b1 , . . . , bn are false, then X is

false.
If some condition bi is satisfied then X is true, unless some inhibitor prevents it. The probability of the inhibitor for bi being
active is denoted by qi . If bi is the only satisfied condition, it
follows that P(X = true) = 1 qi .
The mechanism that inhibits one satisfied condition bi from causing X to be true is independent of the mechanism that inhibits
another satisfied condition bj (i 6= j) from causing X to be true.
The causing conditions combine disjunctively, meaning that if a
number of conditions are satisfied then X is true, unless all the
corresponding inhibitors are active.
See [14, Section 3.3.2] and [31, Section 4.3.2] for further details.
Note: The boolean arguments of the NoisyOR operator can be arbitrary expressions not just simple variables. For example, to introduce a condition that is always satisfied, specify true as the corresponding boolean expression.

6.8

Generating tables

Normally, the user doesnt need to worry about generating tables from their
corresponding models. This is automatically taken care of by the compilation, propagation, and reset-inference-engine operations (by calling the
functions described below).
However, it may sometimes be desirable to generate a single table from its
model (for example, when deciding how to split a continuous range into
subintervals). This is done using the following function.
x

h status t h node generate table (h node t node)

Generate the table of node from its model (a missing model is a usage error).
Since real-valued function nodes do not have tables, node must not be a realvalued function node.
The table will only be generated if the inference engine thinks it is necessary. To be precise: If one of the operations in the list below has been
performed since the most recent generation (if any) of the table, then the
table is [re]generated.
The model of node is new, or one of its expressions is new.
The number of samples per interval for the model of node has changed;
see h model set number of samples per interval(98) below.
95

 A state label (of a labeled node), a state value, the number of states,
or the subtype of node (if node is a discrete node) or one of its discrete
parents has changed.
The value of a parent linked to node by a functional link has (or might
have) changed provided the parent is used in the model of node.6
This includes the case where the parent has been replaced by another
parent using h node switch parent(34) .
If the value cant be computed, the table wont be generated (i.e., the
h node generate table operation fails).
A parent of node has been removed (provided the parent was used in
the model of node).7
If the operation fails, the contents of the table will be undefined. If a log-file
has been specified (see h domain set log file(109) ), then information about
the computations (including reasons for failures) is written to the log-file.
Experience and fading tables (see Chapter 11) are not affected by h node
generate table.
Generation of tables is usually a static operation. That is, tables can be generated once prior to compilation and inference. But this is not always the
case:8 If there are functional trails between nodes that are not real-valued
function nodes, then tables might have to be generated during inference
see Section 10.2.
h status t h domain generate tables (h domain t domain)

Generate tables for all relevant nodes of domain. This is done by calling h
node generate table for all nodes (except real-valued function nodes) having
a model. Hence, the description of that function also applies here.
The operation is aborted if table generation fails for some node. This implies
that some tables may have been successfully generated, some may not have
been generated at all, and one table has been only partially generated.
The following function is identical to the above function, except that it operates on classes instead of domains.
h status t h class generate tables (h class t class)

Generate tables for all relevant nodes of class. See the above description of
h domain generate tables for further details.
6

Parents that only appear within probability and aggregate expressions are ignored, because their values are not needed for the purpose of table generation.
7
Adding a parent, however, will not cause the table to be generated (because the contents
of the table would not change).
8
This changed in HUGIN 7.7.

96

As mentioned above, information about the computations performed when

generating a table is written to the log-file. For classes, a log-file is specified
using the following function.
h status t h class set log file (h class t class, FILE log file)

Set the file to be used for logging by subsequent HUGIN API operations that
apply to class. (Currently, only table generation information is written to
log-files for classes.)
If log file is NULL, no log will be produced. If log file is not NULL, it must
be a stdio text file, opened for writing (or appending). Writing is done
sequentially (i.e., no seeking is done).
See also Section 7.4.

6.9

How the computations are done

In this section, we consider the computations that must be performed when

a node table is generated from a model.
The node table of a discrete function node only contains the node itself. In
this case, the values of the parents can be used in the expressions of the
model. These values are obtained as explained in Section 9.7. Then generation of the table proceeds as in the case of chance (and decision) nodes.
Otherwise, for each configuration of the (discrete) parents, we must compute either an appropriate value or a (conditional) probability distribution.
[The first case is used for utility nodes and for deterministic relationships
see below.] In either case, the expression corresponding to the given configuration is extracted from the model and evaluated assuming that the parents
have been assigned values as follows:
If the parent is a real-valued function node, the value is obtained from
h node get value(128) .
If the parent is a non-interval discrete node, the value is simply the
truth value, label, or state value associated with the state index of the
parent in the given configuration.
If the parent is an interval node, then the expression is evaluated using
many values in the interval corresponding to the state index of the parent in the given configuration. This is explained in more detail below.

Computations for interval parents

Assume, for simplicity, that we have a discrete child node with one interval
parent (more than one interval parent is a trivial generalization of the simple
case).
97

For a given interval of the parent (i.e., for a specific parent state configuration), we compute many probability distributions for the child, each distribution being obtained by instantiating the parent to a value in the interval
under consideration.9 The average of these distributions is used as the conditional probability distribution for the child given the parent is in the interval state considered. (For this scheme to work well, the intervals should be
chosen such that the discretised distributions corresponding to the chosen
points in the parent interval are not too different from each other.)
By default, 25 values are taken within each bounded interval of an interval
parent: The interval is divided into 25 subintervals, and the midpoints of
these subintervals are then used in the computations. A large number of
values gives high accuracy, and a small number of values results in fast
computations. The number of values used can be changed by the following
function.
h status t h model set number of samples per interval
(h model t model, size t count)

Specify that count values should be sampled from each bounded interval of
an interval parent when generating a table from model.
Note that this has no effect if the owner of model is a function node (of
any kind).
h count t h model get number of samples per interval
(h model t model)

Retrieve the count indicating the number of samples that would be used if a
table were to be generated from model now.

Deterministic relationships
If the type of the expression for the parent state configuration under consideration is not distribution, then we have a deterministic relationship.
The expression must then evaluate to something that matches one of the
states of the child node. For labeled, boolean, and numbered nodes, the
value must match exactly one of the state values or labels. For interval
nodes, the value must belong to one of the intervals represented by the
states of the child node.
If one or more of the parents are of interval subtype, then a number of
samples (25 by default) within each (bounded) interval will be generated.
Each of these samples will result in a degenerate distribution (i.e., all
9

For semi-infinite intervals, only one value is used. This value is chosen to be close to the
finite endpoint. Intervals that are infinite in both directions are discouraged the behavior
is unspecified.

98

probability mass will be assigned to a single state) for the child node. The
final distribution assigned to the child node is the average over all generated
distributions. This amounts to counting the number of times a given child
state appears when applying the deterministic relationship to the generated
samples.
If all samples within a given parent state interval map to the same child
state, then the resulting child distribution is independent of the number of
samples generated. It is recommended that the intervals be chosen such
that this is the case.
If this is not feasible, then the number of samples generated should be large
in order to compensate for the sampling errors. Typically, some of the child
states will have a frequency count one higher (or lower) than the ideal
count.
Example 6.5 Let X be an interval node having [0, 1) as one of its states (intervals).
Let Y be a child of X having [0, 1), [1, 2), and [2, 3) as some of its states. Assume
that Y is specified through the deterministic relation Y = 3X. If 25 samples for X
are taken within the interval [0, 1), then 8, 9, and 8 of the computed values will
fall in the intervals [0, 1), [1, 2), and [2, 3) of Y, respectively. Ideally, the frequency
counts should be the same, resulting in a uniform distribution over the three interval states.

99

100

Chapter 7

Compilation
Before a belief network or a LIMID can be used for inference, it must be
compiled.
This chapter describes functions to compile domains, control triangulations,
and to perform other related tasks, such as approximation and compression.

7.1

What is compilation?

The compilation process involves the following sequence of steps:1

(1) The real-valued (but not the discrete) function nodes are removed.
(2) The links between (discrete) function nodes and their parents are removed.
(3) The network is converted into its moral graph: The parents of each
node are married (i.e., links are added between them), and the directions of the links are dropped.
(4) The utility nodes are removed.
(5) The graph is triangulated. (This is described in detail below.)
(6) The cliques (maximal complete sets) of the triangulated graph are
identified, and the collection of cliques is organized as a tree (with
the cliques forming the nodes of the tree). Such a tree is called a junction tree. If the original network is disconnected, there will be a tree
for each connected component.
1
The network presented for compilation is not actually modified by these steps. The
compilation process should be thought of as working on a copy of the original network.

101

(7) Finally, potentials are associated with the cliques and the links (the
separators) of each junction tree. These potentials are initialized from
the evidence and the conditional probability tables (and the policies
and the utility tables in the case of LIMIDs), using a sum-propagation
(see Section 10.1).
All steps, except the triangulation step, are quite straightforward. The triangulation problem is known to be NP-hard for all reasonable criteria of
optimality, so (especially for large networks) finding the optimal solution
is not always feasible. The HUGIN API provides several methods for finding triangulations: five heuristic methods based on local cost measures, and
a combined exact/heuristic method capable of minimizing the storage requirements (i.e., the sum of state space sizes) of the cliques of the triangulated graph, if sufficient computational resources are available.
Alternatively, a triangulation can be specified in the form of an elimination
sequence.
An elimination sequence is an ordered list containing each node of the graph
exactly once. An elimination sequence hv1 , . . . , vn i generates a triangulated
graph from an undirected graph as follows: Complete the set of neighbors
of v1 in the graph (i.e., for each pair of unconnected neighbors, add a fillin edge between them). Then eliminate v1 from the graph (i.e., delete v1
and edges incident to v1 ). Repeat this process for all nodes of the graph in
the specified order. The input graph with the set of generated fill-in edges
included is a triangulated graph.
The elimination sequence can be chosen arbitrarily, except for belief networks with both discrete and continuous nodes.
In order to ensure correct inference, the theory of CG belief networks (see
[8, 20, 23]) requires the continuous nodes to be eliminated before the discrete nodes.
Let denote the set of discrete nodes, and let denote the set of continuous
nodes. A valid elimination sequence must contain the nodes of (in any
order) followed by the nodes of (in any order).
Let x, y . It is well-known that, for any valid elimination sequence, the
following must hold for the corresponding triangulated graph: If between x
and y there exists a path lying entirely in (except for the end-points), then
x and y are connected. If x and y are not connected in the moral graph,
we say that x and y are connected by a necessary fill-in edge. Conversely,
it can be shown that a triangulated graph with this property has a valid
elimination sequence.
Let G be the moral graph extended with all necessary fill-in edges. The
neighbors of a connected component of G[ ] form a complete separator of G
(unless there is exactly one connected component having all nodes of as
neighbors). A maximal subgraph that does not have a complete separator
102

is called a prime component [30]. We note that a triangulation formed by

the union of triangulations of the prime components of G has the property
described above.
We can now further specify the triangulation step. The input to this step is
the moral graph (after removal of utility nodes).
Add all necessary fill-in links.
Find the prime components of the graph.
Triangulate the prime components. The union of these triangulations
constitutes the triangulation of the input graph.
Generate an elimination sequence for the triangulation.
See [8, 13, 14, 20] for further details on the compilation process.

7.2

Compilation
h status t h domain compile (h domain t domain)

Compile domain, using the default triangulation method (unless domain is

already triangulated see Section 7.3); domain must contain at least one
discrete or continuous node.
It is considered an error, if domain is already compiled.
The junction trees are initialized using the sum-propagation operation, incorporating all available evidence in normal mode. If this propagation
fails, the compilation also fails.
The compilation process can use large amounts of memory (and time), depending on the size and structure of the network, and the choice of triangulation method. The application should be prepared to handle out-of-memory
conditions.
h boolean t h domain is compiled (h domain t domain)

Test whether domain is compiled.

7.3

Triangulation

The choice of triangulation method in the compilation process can have a

huge impact on the size of the compiled domain, especially if the network
is large. If the default triangulation method used by h domain compile(103)
does not give a good result, another option is available: The network may
103

be triangulated in advance (before calling h domain compile). The HUGIN

API provides a choice between several built-in triangulation methods, or,
alternatively, a triangulation can be specified in the form of an elimination
sequence.
The HUGIN API defines the enumeration type h triangulation method t.
This type contains the following six constants, denoting the built-in triangulation methods: h tm clique size, h tm clique weight, h tm fill in size, h
tm fill in weight, h tm best greedy, and h tm total weight. The first four of
these methods are simple elimination based heuristics, the fifth method is
a combination of the first four methods, and the last method is a combined
exact/heuristic method capable of producing an optimal triangulation, if
sufficient computational resources (primarily storage) are available.
The elimination based heuristics follow a common scheme: Nodes are eliminated sequentially in a greedy manner (that is, with no look-ahead) from
the prime component being triangulated. If the current graph has a node
with all its (uneliminated) neighbors forming a complete set, that node is
eliminated next (this is optimal with respect to minimizing the size of the
largest clique of the final triangulated graph). If no such node exists, a node
with a best score according to the selected heuristic is chosen.
Let C(v) denote the set comprised of v and its (uneliminated) neighbors.
The elimination based heuristics, implemented in HUGIN, define a score
based on C(v) for each node v (and with best defined as minimum).
The scores defined by the four basic heuristics are:
h tm clique size The score is equal to the cardinality of C(v).
h tm clique weight The score is s + sm(m + 3)/2, where s is the product
of the number of states of the discrete nodes in C(v), and m is the
number of continuous nodes in C(v). This score is equal to the number
of data elements stored in a table holding C(v).2
h tm fill in size The score is equal to the number of fill-in edges needed to
complete C(v).
h tm fill in weight The score is equal to the sum of the weights of the fillin edges needed to complete C(v), where the weight of an edge is
defined as the product of the number of states of the nodes connected
by the edge (in this context, continuous nodes are equivalent to discrete nodes with only one state).
2

This is not always proportional to the actual storage requirements (measured in bytes)
of the table. This is because the data elements can be of different types: The data elements
associated with state configurations of discrete nodes are of type h number t, while the data
elements associated with continuous nodes (such as mean and covariance values) are of type
h double t. In a single-precision version of the HUGIN API, h number t is a 4-byte quantity,
but in a double-precision version, it is an 8-byte quantity. In both versions, h double t is an
8-byte quantity.

104

It is possible for these heuristics to produce a non-minimal triangulation (a

triangulation is minimal if no proper subset of the fill-in edges produces a
triangulated graph). In order to obtain a minimal triangulation, a maximal
set of redundant fill-in edges is removed from the triangulation. This is done
using a variant of an algorithm by Kjrulff [16].
The h tm best greedy triangulation method combines the results of the four
basic heuristics: It tries all of the heuristics and uses the best result. (Here,
best result refers to the triangulation with the smallest sum of clique
weights as defined by the h tm clique weight heuristic.) Notice that different heuristics can produce the best result on different prime components.
The h tm best greedy heuristic is the method used by h domain compile(103) ,
if the domain being compiled has not been triangulated in advance.
The heuristic triangulation methods are very fast, but sometimes the generated triangulations are quite bad. As an alternative, the HUGIN API provides the h tm total weight triangulation method. This method can produce
an optimal triangulation, if sufficient computational resources are available.
The method considers a triangulation to be optimal, if it is minimal and
the sum of clique weights (as defined by the h tm clique weight heuristic) is
minimum.
For some large networks, use of the h tm total weight triangulation method
has improved time and space complexity of inference by an order of magnitude (sometimes even more), compared to the heuristic methods described
above.
The h tm total weight triangulation algorithm can be outlined as follows:
First, all minimal separators of the prime component being triangulated are
identified (using an algorithm by Berry et al [2]). From this set of minimal separators, an initial triangulation is found using greedy selection of
separators until the component is triangulated. The cost of this triangulation is then used as an upper bound on the cost of optimal triangulations,
and all separators that are too expensive relative to this upper bound are
discarded. Then the prime component is split using each of the remaining
separators. The pieces resulting from such a split are called fragments (cf.
Shoikhet and Geiger [34]; another term commonly used is 1-block). Every
(minimal) triangulation of a fragment can be decomposed into a clique and
corresponding triangulated subfragments. Using this fact, optimal triangulations are found for all fragments. (The clique generation process of this
search makes use of a characterization of cliques in minimal triangulations
given by Bouchitte and Todinca [3].) Finally, an optimal triangulation of
the prime component is identified (by considering all possible splits of the
component using minimal separators).
Initial triangulation. Instead of using greedy selection of minimal separators
to find the initial triangulation, the user can specify a triangulation in the
105

form of an elimination sequence. If a good triangulation is already known,

then this can be used to speed up the triangulation process significantly.
h status t h domain set initial triangulation
(h domain t domain, h node t order)

Specify an intial triangulation to be used by the h tm total weight triangulation method. The triangulation is specified in the form of the elimination
sequence order (a NULL-terminated list of nodes containing each discrete
and each continuous node of domain exactly once, and continuous nodes
must precede discrete nodes see Section 7.1).
If NULL is specified as the initial triangulation, then any previously specified
initial triangulation is removed. The initial triangulation is also removed if a
new (discrete or continuous) node is created within domain, or an existing
(discrete or continuous) node is deleted.
The number of separators. Some prime components have more minimal separators than the memory of a typical computer can hold. In order to handle such components, an upper bound on the number of separators can be
specified: If the search for minimal separators determines that more than
the specified maximum number of separators exist, then the component is
split using one of the separators already found.3 The fragments obtained
are then recursively triangulated.
Experience suggests that 100 000 is a good number to use as an upper bound
on the number of minimal separators.
h status t h domain set max number of separators
(h domain t domain, size t count)

Specify count as the maximum number of minimal separators to generate

when using the h tm total weight method for triangulating domain. If count
is zero, then the bound is set to unlimited, which is also the default value.
h count t h domain get max number of separators
(h domain t domain)

Retrieve the current setting for domain of the maximum number of separators to generate for the h tm total weight triangulation method. If an error
occurs, a negative number is returned.
The size of separators. The algorithm by Berry et al [2] for finding minimal
separators generate new separators from separators already found. Large
separators are only useful to ensure that all separators are eventually found.
If large separators are ignored (discarded), less storage is consumed, and
3

The separator is selected using a heuristic method that considers the cost of the separator
and the size of the largest fragment generated, when the component is split using the separator. The heuristic method used for this selection may change in a future version of the
HUGIN API.

106

larger networks can be handled (without the need for splitting prime components into smaller graphs). The drawback is that finding all relevant separators (i.e., the small separators that are useful for generating triangulations) is no longer assured.
x

h status t h domain set max separator size

(h domain t domain, size t size)
Specify size as the maximum size of minimal separators to keep when using
the h tm total weight method for triangulating domain. If size is zero, then
the bound is set to unlimited, which is also the default value.
This can be used in combination with h domain set max number of separators(106) .
Because some useful separators might not be found when using the maxseparator-size parameter, an initial triangulation must be specified. All
minimal separators are exctracted from this triangulation (with redundant
fill-in edges removed) and are used (i.e., considered for triangulation regardless of their sizes) to ensure that a triangulation can always be found.

h count t h domain get max separator size (h domain t domain)

Retrieve the current setting for domain of the maximum size of separators
to keep for the h tm total weight triangulation method. If an error occurs, a
negative number is returned.
The following is the actual triangulation function: It triangulates domain
using the specified triangulation method.

h status t h domain triangulate

(h domain t domain, h triangulation method t tm)
Perform triangulation of domain using triangulation method tm. It is considered an error, if domain is already triangulated.
As mentioned above, it is also possible to supply a triangulation explicitly
through an elimination sequence. This is convenient if a better triangulation
is available from other sources.

h status t h domain triangulate with order

(h domain t domain, h node t order)
Triangulate domain using the NULL-terminated list order of nodes as elimination sequence. The list must contain each discrete and each continuous
node of domain exactly once, and continuous nodes must precede discrete
nodes see Section 7.1.
It is considered an error, if domain is already triangulated.
The triangulation functions construct the cliques and the junction trees, but
do not allocate storage for the data arrays of the clique and separator tables.
107

This storage (which is the largest part of the storage consumed by a typical
compilation) is allocated by the h domain compile(103) function. However,
the total size of the junction tree tables can be queried before this storage is
allocated see h jt get total size(117) and h jt get total cg size(117) .
x

h boolean t h domain is triangulated (h domain t domain)

Test whether domain is triangulated.

h node t h domain get elimination order (h domain t domain)

Return a NULL-terminated list of nodes in the order used to triangulate domain. If an error is detected (e.g., domain has not been triangulated), NULL
is returned.
The list holding the elimination order is stored within the domain structure.
It must not be deallocated by the application.
As indicated above, it can be a lot of work to find good triangulations. Therefore, it is convenient to store the corresponding elimination orders in separate files for later use. The following function helps in managing such files:
It parses a text file holding a list of node names (separated by spaces, tabs,
newlines, or comments see Section 13.8).

h node t h domain parse nodes

(h domain t domain, h string t file name,
void (error handler) (h location t, h string t, void ),
void data)
This function parses the list of node names stored in the file with name
file name. The node names must identify nodes of domain; it is an error,
if some node cannot be found. If no error is detected, a NULL-terminated
dynamically allocated array holding the nodes is returned. If an error is
detected, NULL is returned.
Note: Only the user application has a reference to the array, so the user
application is responsible for deallocating the array when it is done using it.
The error handler and data arguments are used for error handling. This is
similar to the error handling done by the other parse functions. See Section 13.9 for further information.
The h domain parse nodes function can be used to parse any file containing
a node list (not just node lists representing elimination orders for triangulations). Therefore, for completeness, a similar parse function is provided for
classes:

h node t h class parse nodes

(h class t class, h string t file name,
void (error handler) (h location t, h string t, void ),
void data)
108

7.4

Getting a compilation log

It is possible to get a log of the actions taken by the compilation process (the
elimination order chosen, the fill-in edges created, the cliques, the junction
trees, etc.). Such a log is useful for debugging purposes (e.g., to find out
why the compiled version of the domain became so big).
x

h status t h domain set log file (h domain t domain, FILE log file)
Set the file to be used for logging by subsequent calls to HUGIN API functions.
If log file is NULL, no log will be produced. If log file is not NULL, it must
be a stdio text file, opened for writing (or appending). Writing is done
sequentially (i.e., no seeking is done).
Note that if a log is wanted, and (some of) the nodes (that are mentioned
in the log) have not been assigned names, then names will automatically be
assigned (through calls to the h node get name(42) function).
Example 7.1 The following code fragment illustrates a typical compilation process.
h_domain_t d;
FILE *log;
...
log = fopen ("mydomain.log", "w");
h_domain_set_log_file (d, log);
h_domain_triangulate (d, h_tm_clique_weight);
h_domain_compile (d);
h_domain_set_log_file (d, NULL);
fclose (log);
A file (log) is opened for writing and assigned as log file to domain d. Next, triangulation, using the h tm clique weight heuristic, is performed. Then the domain is
compiled. When the compilation process has completed, the log file is closed. Note
that further writing to the log file (by HUGIN API functions) is prevented by setting
the log file of domain d to NULL.

In addition to the compilation and triangulation functions, the h node generate table(95) , h domain learn structure(173) , h domain learn tables(176) , and
h domain learn class tables(179) functions also use the log file to report errors, warnings, and other information. HUGIN API functions that use h
node generate table internally (for example, the propagation operations call
this function when tables need to be regenerated from their models) also
write to the log file (if it is non-NULL).
109

7.5

Uncompilation
h status t h domain uncompile (h domain t domain)

Remove the data structures of domain produced by the h domain compile(103),

h domain triangulate(107), and h domain triangulate with order(107) functions.
If domain is not compiled or triangulated, nothing is done.
Note that any opaque references to objects within the compiled structure
(e.g., clique and junction tree objects) are invalidated by an uncompile
operation.
Also note that many of the editing functions described in Chapter 2 automatically perform an uncompile operation whenever something is changed
about domain that invalidates the compiled structure. When this happens,
the domain must be recompiled (using h domain compile(103) ) before it can
be used for inference.

7.6

Compression

Most of the memory consumed by a compiled domain is used for storing the
data of the clique and separator tables. Many of the entries of these tables
might be zero, reflecting the fact that these state combinations in the model
are impossible. Zeros in the junction tree tables arise from logical relations
within the model. Logical relations can be caused by deterministic nodes,
approximation, or propagation of evidence. To conserve memory, the data
elements with a value of zero can be removed, thereby making the tables
smaller. This process is called compression.
x

h double t h domain compress (h domain t domain)

Remove the zero entries from the clique and separator tables of the junction
trees of domain.
DBN: domain must not be a DBN runtime domain that has been triangulated
using h domain triangulate dbn(65) .
If a junction tree (probability) table contains continuous nodes, compression
only applies to configurations of discrete nodes: the continuous data associated with a discrete configuration is removed if and only if the probability
associated with the configuration is zero.
In compiled LIMID networks, clique and separator tables come in pairs consisting of a probability and a utility table of identical structure. In this
case, the utility associated with a (discrete) configuration in the utility table
is removed if and only if the probability associated with the same configuration in the probability table is removed.
110

If domain has a memory backup (see h domain save to memory(142) ), it will

be deleted as part of the compression operation.
The compression function returns a measure of the compression achieved.
This measure should be less than 1, indicating that the compressed domain
requires less memory than the uncompressed version. If the measure is
larger than 1, the compressed domain will actually use more memory than
the uncompressed version. This can happen if only a few elements of the
junction tree tables are zero, so that the space savings achieved for the tables
are outweighed by the extra data structures required to support the more
complex table operations needed to do inference in compressed domains.4
If an error occurs, h domain compress returns a negative number.
If a domain has been compressed, and more zeros have been introduced
by new evidence or approximation (Section 7.7), then the domain can be
compressed further to take advantage of the new zero entries.
Note that some operations, such as extracting marginal tables (with more
than one node), cannot be done with a compressed domain. Those operations will fail with the error code h error compressed.
Note also that compression is only possible after compilation is completed.
This means that enough memory to store the uncompressed compiled domain must be available. Compression is maintained in saved domains (when
HUGIN KB files are used), making it possible to use a computer with a large
amount of (virtual) memory to compile and compress a domain and then
load the compressed domain on computers with less memory.
The zero elements in the junction tree tables do not contribute anything to
the beliefs computed by the HUGIN inference engine. Thus, their removal
doesnt change the results of inference. The only effect of compression is to
save memory and to speed up inference.
h boolean t h domain is compressed (h domain t domain)

Test whether domain is compressed.

7.7

Approximation

The discrete part of a clique potential consists of a joint probability distribution over the set of state configurations of the discrete nodes of the clique.
4

Prior to HUGIN API 7.4, a 16-bit integer type was used for table indexes within the data
structures of compressed domains. In version 7.4, this type was changed to a 32-bit integer
type. This allows construction of compressed domains with much larger tables at the cost
of a larger overhead of the data structures needed to support table operations on compressed
tables.

111

The approximation technique implemented in the HUGIN API is based

on the assumption that very small probabilities in this probability distribution reflect (combinations of) events that will hardly ever occur in practice.
Approximation is the process of setting all such near-zero probabilities to
zero. The primary objective of this process is to minimize storage consumption through compression.
It must be emphasized that this approximation technique should only be
used when one is not interested in the probabilities of unlikely states as the
relative error although small in absolute terms for such probabilities
can be huge. Approximation should be used only if all one is interested in is
to identify the most probable state(s) for each node given evidence.
x

h double t h domain approximate

(h domain t domain, h double t )
The effect of this function is as follows: For each clique in domain, a value
is computed such that the sum of all elements less than in (the discrete
part of) the clique table is less than . These elements are then set to 0.
In effect, specifies the maximum probability mass to remove from each
clique.
Approximation can only be applied to (compiled) networks without decision
nodes. [Continuous nodes are allowed, but approximation only applies to
configurations of states of the discrete nodes.]
The type of equilibrium on the junction trees of domain must be sum, incorporating all evidence (if any) specified for domain in normal mode (Section 10.1). This condition holds right after a (successful) compilation, which
is when an approximation is usually performed.
The approximation function returns the probability mass remaining in the
entire domain, letting you know how much precision you have lost. Note
that this is not the same as 1 , as the value is relative to each clique.
Typically, the total amount of probability mass removed will be somewhat
larger than .
If h domain approximate fails, a negative value is returned.
The annihilation of small probabilities within the clique potentials can be
thought of as entering a special kind of evidence. As part of the approximation process, this evidence is propagated throughout the junction trees,
thereby reaching an equilibrium state on all junction trees. The joint probability of the evidence is the value returned by h domain approximate.
An approximation operation should be followed by a compression operation. If not, the approximation will be lost when the inference engine is
reset (which can, e.g., happen as part of a propagation operation when evidence has been retracted and/or some conditional probability tables have
changed).
112

Example 7.2 Example 7.1 can be extended with approximation and compression
as follows.
h_domain_t d;
FILE *log;
...
log = fopen ("mydomain.log", "w");
h_domain_set_log_file (d, log);
h_domain_triangulate (d, h_tm_clique_weight);
h_domain_compile (d);
h_domain_set_log_file (d, NULL);
fclose (log);
h_domain_approximate (d, 1E-8);
h_domain_compress (d);
h_domain_save_as_kb (d, "mydomain.hkb", NULL);
Probability mass of weight up to 108 is removed from each clique of the compiled
domain using approximation. Then the zero elements are removed from the clique
potentials using compression. Finally, the domain is saved as an HKB file (this
is necessary in order to use the compressed domain on another computer with
insufficient memory to create the uncompressed version of the domain).

It is difficult to give a hard-and-fast rule for choosing a good value for

(i.e., one that achieves a high amount of compression and doesnt introduce
unacceptable errors). In general, the ratio between the error introduced by
the approximation process and the joint probability of the evidence obtained
when using the approximated domain should not become too large. If it
does, the evidence should be processed by the unapproximated version. A
threshold value for this ratio should be determined through empirical tests
for the given domain.
See [11] for an empirical analysis of the approximation method.
x

h double t h domain get approximation constant

(h domain t domain)
Return the approximation constant of the most recent (explicit or implicit)
approximation operation. If an error occurs, a negative number is returned.
An implicit approximation takes place when you change some conditional
probability tables of a compressed domain, and then perform a propagation
operation. Since some (discrete) state configurations have been removed
from a compressed domain, the probability mass of the remaining configurations will typically be less than 1; h domain get approximation constant
returns that probability mass.

113

114

Chapter 8

Cliques and Junction Trees

The compilation process creates a secondary structure of the belief network
or LIMID model. This structure, the junction tree, is used for inference.
Actually, since the network may be disconnected,1 there can be more than
one junction tree. In general, we get a forest of junction trees except for
a LIMID, which is compiled into a single junction tree (in order to ensure
correct inference).
The cliques of the triangulated graph form the nodes of the junction trees.
The connections (called separators) between the cliques (i.e., the edges
of the junction trees) are the communication channels used by CollectEvidence and DistributeEvidence (see Chapter 10). Associated with each
clique and separator is a function from the state space of the clique/separator
to the set of (nonnegative) real numbers this is called the (probability) potential, and it is represented as a table (Chapter 5). If the input to the compilation process contains utilities, there will be an additional potential associated with the cliques and the separators: a utility potential which is similar
to the probability potential (except that the numbers may be negative).
The triangulation process constructs the cliques and the junction trees, but
does not allocate storage for the data arrays of the clique and separator
tables. This storage (which is the largest part of the storage consumed by
a typical compilation) is allocated by the h domain compile(103) function.
However, the total size of the junction tree tables can be queried before
this storage is allocated. This is useful for evaluating the quality of the
triangulation.
The HUGIN API provides functions to access the junction forest and to traverse the trees of the forest.

1
Actually, since functional links are ignored by the compilation process, the network
can be connected and generate more than one junction tree.

115

8.1

Types

We introduce the opaque pointer types h clique t and h junction tree t.

They represent clique and junction tree objects, respectively.

8.2

Junction trees

The HUGIN API provides a pair of functions to access the junction trees of a
triangulated domain.
x

h junction tree t h domain get first junction tree

(h domain t domain)
Return the first junction tree in the list of junction trees of domain. If an error is detected (e.g., domain is not triangulated), a NULL pointer is returned.

h junction tree t h jt get next (h junction tree t jt)

Return the successor of junction tree jt. If there is no successor, NULL is
returned. If an error is detected (i.e., if jt is NULL), NULL is returned.
Another way to access junction trees is provided by the h clique get junction
tree and h node get junction tree functions.

h junction tree t h clique get junction tree (h clique t clique)

Return the junction tree to which clique belongs. If an error is detected,
NULL is returned.

h junction tree t h node get junction tree (h node t node)

Return the junction tree to which node belongs (node must not be a function
node). If an error is detected, NULL is returned.
As explained in Section 7.1, utility (and function) nodes are not present in
the junction trees. But because only one junction tree is constructed for a
LIMID network, h node get junction tree returns that junction tree if node is
a utility node.
We also provide a function to access the collection of cliques comprising a
given junction tree.

h clique t h jt get cliques (h junction tree t jt)

Return a NULL-terminated list of the cliques that form the set of vertices of
junction tree jt. If an error is detected, a NULL pointer is returned.
The storage holding the list of cliques returned by h jt get cliques is owned
by the junction tree object, and must therefore not be deallocated by the
application.
116

h clique t h jt get root (h junction tree t jt)

Return the root of junction tree jt. If the junction tree is undirected (which
it is unless there are continuous nodes involved), this is just an arbitrarily
selected clique. If the junction tree is directed, a strong root (see [8, 20, 23])
is returned (there may be more than one of those). If an error is detected,
NULL is returned.
size t h jt get total size (h junction tree t jt)

Return the total size (i.e., the total number of discrete configurations) of all
clique and separator tables associated with junction tree jt.
Each discrete table configuration has a numeric quantity of type h number t
associated with it. In a single-precision version of the HUGIN API, this is a
4-byte quantity. In a double-precision version, this is an 8-byte quantity.
Note that both probability and utility tables are counted (that is, the discrete
clique and separator configurations are counted twice if there are utility
potentials in the junction tree).
If an error occurs (e.g., the total size of all tables exceeds the maximum
value of the size t type), (size t) 1 is returned.
size t h jt get total cg size (h junction tree t jt)

Return the total CG size of all clique and separator tables associated with
junction tree jt. This counts the total number of CG data elements of all
tables. Each such data element occupies 8 bytes.
If the junction tree contains no CG nodes, the tables contain no CG data. In
this case (only), the function returns 0.
If an error occurs, (size t) 1 is returned.

8.3

Cliques

Each clique corresponds to a maximal complete set of nodes in the triangulated graph. The members of such a set can be retrieved from the corresponding clique object by the following function.
x

h node t h clique get members (h clique t clique)

Return a NULL-terminated list of the nodes comprising the members of clique.
If an error is detected, NULL is returned.
The storage holding the list of nodes is the actual member list stored within
clique, and must therefore not be freed by the application.
117

h clique t h clique get neighbors (h clique t clique)

Return a NULL-terminated list of cliques containing the neighbors of clique

in the junction tree to which clique belongs. If an error is detected, NULL is
returned.
The storage used for holding the clique list returned by h clique get neighbors is owned by clique, and must therefore not be freed by the application.

8.4

Traversal of junction trees

The h jt get root(117) and h clique get neighbors(118) functions can be used
to traverse a junction tree in a recursive fashion.
Example 8.1 The following code outlines the structure of the DistributeEvidence
function used by the propagation algorithm (see [13] for further details).
void distribute_evidence (h_clique_t self, h_clique_t parent)
{
h_clique_t *neighbors = h_clique_get_neighbors (self);
if (parent != 0)
/* absorb from parent */ ;
for (h_clique_t *n = neighbors; *n != 0; n++)
if (*n != parent)
distribute_evidence (*n, self);
}
...
{
h_junction_tree_t jt;
...
distribute_evidence (h_jt_get_root (jt), 0);
...
}
The parent argument of distribute evidence indicates the origin of the invocation;
this is used to avoid calling the caller.

118

Chapter 9

Evidence and Beliefs

The first step of the inference process is to make the inference engine aware
of the evidence. This is called entering the evidence, and it can be done before
or after the compilation step (but note that the compilation process uses
the available evidence to initialize the junction tree potentials). Moreover,
an uncompile operation does not remove any already entered evidence
(this is worth noting because many HUGIN API functions perform implicit
uncompile operations).
Typically, an item of evidence has the form of a statement specifying the
state (or value) of a variable. This type of evidence is entered to the HUGIN
inference engine by the h node select state(121) function for discrete nodes
and by the h node enter value(122) function for continuous nodes. However,
for discrete chance and function nodes, more general items of evidence,
called likelihood or multi-state evidence, can be specified: This type of
evidence is entered using the h node enter finding(122) function.
This chapter explains how to enter and retract evidence, how to determine
independence properties induced by evidence and network structure, how
to retrieve beliefs and expected utilities, how to compute values of function
nodes, how to examine evidence, and how to save evidence as a case file for
later use.
Chapter 10 explains how to propagate evidence in order to compute updated beliefs and other results.

9.1
9.1.1

Evidence
Discrete evidence

Associated with each discrete node in a HUGIN domain model is a function

that assigns a nonnegative real number to each state of the node. We sometimes refer to such a function as an evidence potential or a finding vector.
119

Initially, before any evidence has been entered, all finding vectors consist of
1-elements only. Such evidence is termed vacuous.
If the finding vector for a node has exactly one positive element, the node
is said to be instantiated. The function h node select state(121) instantiates a
node to a specific state (using 1 as the finding value of the specified state).
In general, specifying 0 as the finding value of a state declares the state to
be impossible. All finding vectors must have at least one positive element.
If not, inference will fail with an impossible evidence error code: h error
inconsistency or underflow(136) .
If a finding vector has two (or more) 1-elements, and the remaining elements are 0, we call the finding vector a multi-state finding.
If a finding vector has at least one element that is 6= 0 and 6= 1, the finding
vector is called a likelihood. The following examples illustrate the use of
likelihoods.
Example 9.1 Let A be the node that we wish to enter likelihood evidence for. Now,
suppose we add a new node B as a child of A and specify the conditional probability
table P(B|A) as follows:

b1
b2

a1

a2

0.3
0.7

0.4
0.6

Then, entering the observation B = b1 in the modified network is equivalent to

entering the likelihood (0.3, 0.4) for A in the original network.
This feature can be used for inexact observations. Suppose A represents something
that we cannot observe with 100% certainty, and B represents our observation of A
(such that state bi of B corresponds to state ai of A). If there is a 10% risk of
making a wrong observation, then P(B|A) would be:

b1
b2

a1

a2

0.9
0.1

0.1
0.9

If B is part of the network, then we would enter either B = b1 or B = b2 according

to our actual observation. If B is not part of the network, we would instead enter
either (0.9, 0.1) or (0.1, 0.9) as likelihood for A.
Example 9.2 Suppose we want to make inference pretending that some root
node has some other prior distribution than the specified P(A). This can be done
by specifying a likelihood equal to the quotient of the desired prior and the original
prior. (This trick, of course, only works when division by zero is not involved.)

An instantiated node is said to have hard evidence. All other types of (nonvacuous) evidence are called soft evidence.
120

9.1.2

Continuous evidence

Evidence for continuous nodes always take the form of a statement that a
node is known to have a specific value. Such evidence is entered using the
h node enter value(122) function.
This type of evidence is an example of hard evidence.

9.1.3

Evidence in LIMIDs

Decision nodes can only have hard evidence (and the finding value must
be 1). In addition, a chance node with evidence must not have a decision node without evidence as ancestor in the network obtained by ignoring information links (i.e., links pointing at decision nodes). Such an evidence scenario would amount to observing the consequences of a decision
before the decision is made, and an attempt to perform inference given
such evidence fails with an invalid evidence error code: h error invalid
evidence(137) .

9.2

Entering evidence

The functions described in this section can be used to enter evidence for
a given set of nodes (one node at a time). It is also possible to load the
evidence for all nodes at once, when the evidence is stored in a case file
(see h domain parse case(132) ) or as a case in main memory (see h domain
enter case(168) ).
The following function handles evidence taking the form of instantiations of
discrete variables.
x

h status t h node select state (h node t node, size t state)

Select state of node (which must be a discrete node). This is equivalent to
specifying the finding value 1 for state and 0 for all other states (see also
h node enter finding below).
The enumeration of the states of a node follows traditional C conventions;
i.e., the first state has index 0, the second state has index 1, etc. So, if node
has n states, then state must be a nonnegative integer smaller than n.
Example 9.3 The following code
h_domain_t d = h_kb_load_domain ("mydomain.hkb", NULL);
h_node_t n = h_domain_get_node_by_name (d, "input");
h_node_select_state (n, 0);
...
loads a domain and enters the observation that node input is in state 0.

121

If the evidence is not a simple instantiation, then the function h node enter
finding should be called, once for each state of the node, specifying the
finding value for the state.
h status t h node enter finding
(h node t node, size t state, h number t value)

Specify value as the finding value for state of node (which can be any discrete
node1 ). The finding value must be nonnegative, and state must specify a
valid state of node.
If you have several independent observations to be presented as likelihoods
to HUGIN for the same node, you have to multiply them yourself; each call
to h node enter finding overrides the previous finding value stored for the
indicated state. The h node get entered finding(129) function can be conveniently used for the accumulation of a set of likelihoods.
To specify evidence for a continuous node, the following function must be
used.
h status t h node enter value (h node t node, h double t value)

Specify that the continuous node node has the value value.
Note that inference is not automatically performed when evidence is entered
(not even when the domain is compiled). To get the updated beliefs, you
must explicitly call a propagation function (see Section 10.2).

9.3

Retracting evidence

If an already entered observation is found to be invalid, it can be retracted

by the following function.
x

h status t h node retract findings (h node t node)

Retract all findings for node. [If node is discrete, this is equivalent to setting
the finding value to 1 for all states of node.]

h status t h domain retract findings (h domain t domain)

Retract findings for all nodes of domain. This is useful when, e.g., a new set
of observations should be entered; see also h domain initialize(142) .
DBN: Evidence that has been moved out of the time window by h domain
move dbn window(66) will not be retracted.
In addition to h node retract findings, h domain retract findings, and h domain initialize (and deletion of domains and nodes, of course), the h node
1
Because evidence for a decision node must always be an instantiation, it is usually better
to use h node select state to specify evidence for decisions.

122

set number of states(38) function deletes evidence when the number of states
of a (discrete) node is changed.
Example 9.4 The code
...
d = h_kb_load_domain ("mydomain.hkb", NULL);
n = h_domain_get_node_by_name (d, "input");
h_node_select_state (n, 0);
...
h_node_retract_findings (n);
enters the observation that the discrete node input is in state 0; later, that observation is retracted, returning the node input to its initial status.

9.4

Determining independence properties

Independence properties in belief networks are often determined using the

concept of d-separation [8, 9, 14, 18, 19, 22, 31, 32, 36]:
Given three sets of nodes: A, B, and S. A trail (a sequence of
nodes that are connected by arcs, of any directionality) from a
node a A to a node b B in an acyclic directed graph is said
to be blocked by S, if there is a node n such that either
n S and arcs of do not meet head-to-head at n, or
n and all its descendants are not in S, and arcs of meet
head-to-head at n.
A trail that is not blocked by S is said to be active.
If all trails between A and B are blocked, then A and B are said
to be d-separated given S. We denote this as AB|S.
Notice that, in the above definition, the sets A, B, and S are not assumed to
be disjoint. Instead, we have, for x S and any y, that {x}{y}|S.
Intuitively, when we have beliefs computed conditional on evidence on some
set E of nodes, and we receive new evidence on a set E 0 of nodes, then
nodes X such that {X}E 0 |E will not be affected by the new evidence (that
is, after propagation of E 0 , the beliefs of X will be unchanged).
Lauritzen et al [22] describes a different (but equivalent) criterion for determining the independence properties in belief networks.
Given three sets of nodes: A, B, and S. Construct the induced
subgraph consisting of A B C and their ancestors, and then
form the moral graph corresponding to this subgraph. If S separates A \ S and B \ S in this moral graph, we have AB|S.
123

This criterion forms the basis of the algorithm in the HUGIN API for determining independence properties.
The following functions determine the (maximal) sets of nodes that are
d-connected to (respectively d-separated from) the specified source nodes
given evidence nodes.
x

h node t h domain get d connected nodes

(h domain t domain, h node t source, h node t hard,
h node t soft)

h node t h domain get d separated nodes

(h domain t domain, h node t source, h node t hard,
h node t soft)
Return the maximal set of nodes that is d-connected to (respectively d-separated from) the source nodes given the evidence nodes.
The evidence nodes are specified using the hard and soft node lists: hard
must contain the nodes with hard evidence, and soft must contain the nodes
with soft evidence (see Section 9.1). These functions treat nodes with soft
evidence as having no evidence themselves but instead as having (imaginary) instantiated child nodes see Example 9.1.
The d-separation criterions (as stated above) only handle chance nodes. In
order to handle non-chance nodes, the algorithm is modified as follows:
In LIMIDs, links pointing at instantiated decision nodes are ignored, but
links pointing at uninstantiated decision nodes are not. Also, the evidence
specified in the hard and soft node lists must form a valid evidence scenario
(see Section 9.1.3), and, moreover, adding the source nodes to the set of
nodes with hard evidence must also yield a valid evidence scenario. (An
invalid evidence error code is set if these constraints are not satisfied.
The constraints also imply that only chance, decision, and discrete function
nodes can appear in the source, hard, and soft node lists.)
For functional links, information must follow the specified directions of
the links. This implies that links entering function nodes and exiting realvalued function nodes are ignored when the ancestral graph is constructed
(that is, the ancestral graph does not contain real-valued function nodes).
However, real-valued function (and utility) nodes can be reached if they
are descendants of a reachable node in the ancestral graph. So, the lists
of d-connected and d-separated nodes may contain nodes of all categories
(except instance nodes).
h domain get d connected nodes and h domain get d separated nodes return
their results in a permanently allocated node list stored inside domain, so the
application must not free the list. The list is overwritten by each (successful)
call to any of the functions.
If an error occurs, the functions return NULL.
124

9.5

Retrieving beliefs

When the domain has been compiled (see Chapter 7) and the evidence propagated (see Chapter 10), the calculated beliefs can be retrieved using the
functions described below.
x

h number t h node get belief (h node t node, size t state)

The belief in state state of the discrete node node is returned. If an error
is detected (e.g., state is an invalid state, or node is not a discrete node), a
negative number is returned.
Note that if evidence has been entered since the most recent propagation,
the beliefs returned by this function may not be up-to-date.
The beliefs in the states of a node usually form a probability distribution for
the node. However, other possibilities exist, determined by the propagation
method used; see Chapter 10 for details.
Example 9.5 A sample use of a domain could be
h_domain_t d;
h_node_t n;
int i, k;
...
n = h_domain_get_node_by_name (d, "input");
h_node_select_state (node, 0);
h_domain_propagate (d, h_equilibrium_sum, h_mode_normal);
n = h_domain_get_node_by_name (d, "output");
k = h_node_get_number_of_states (n);
for (i = 0; i < k; i++)
printf ("P(output=%d|input=0) = %g\n", i,
h_node_get_belief (n, i));
...
This code enters the observation that node input is in state 0. This observation
is then propagated to the remaining nodes, using h domain propagate(135) . Finally,
the revised beliefs (the conditional probabilities given the observation) for node
output are displayed.

For continuous nodes, the beliefs computed take the form of the mean and
variance of the distribution of the node given the evidence.
x

h double t h node get mean (h node t node)

Return the mean of the marginal distribution of the continuous node node.

h double t h node get variance (h node t node)

Return the variance of the marginal distribution of the continuous node
node.
125

The marginal distribution of node is not necessarily a Gaussian distribution.

In general, it is a mixture of several Gaussians. See the description of the
h node get distribution(126) function for instructions on how to access the
individual components of the mixture.
Sometimes, the joint distribution over a set of nodes is desired:
h table t h domain get marginal
(h domain t domain, h node t nodes)

Compute the marginal table for the specified list nodes of nodes with respect
to the (imaginary) joint potential, determined by the current potentials on
the junction tree(s) of domain. The nodes must be distinct discrete or continuous nodes, and they must belong to domain.2 If the nodes list contains
continuous nodes, they must be last in the list. This operation is not allowed
on compressed domains. If an error occurs, a NULL pointer is returned.
The fact that the marginal is computed based on the current junction tree
potentials implies that the equilibrium and evidence incorporation mode
(see Section 10.1) for the marginal will be as specified in the propagation
that produced the current junction tree potentials.
If the nodes list contains continuous nodes, the marginal will in general be
a so-called weak marginal [8, 20, 23]. This means that only the means and
the (co)variances are computed, not the full distribution. In other words,
the marginal is not necessarily a multi-variate normal distribution with the
indicated means and (co)variances (in general, it is a mixture of such distributions). Also note that if the discrete probability is zero, then the mean and
(co)variances are essentially random numbers (the inference engine doesnt
bother computing zero components of a distribution).
The table returned by h domain get marginal is owned by the application,
and it is the responsibility of the application to deallocate it (using h table
delete(73) ) after use.
See Chapter 5 for information on how to manipulate h table t objects.
h table t h node get distribution (h node t node)

This function computes the distribution for the CG node node. No value
must be propagated for node. If an error occurs, a NULL pointer is returned.
The distribution for a CG node is in general a mixture of several Gaussian
distributions. What h node get distribution really computes is a joint distribution for node and a set of discrete nodes. The set of discrete nodes is
chosen such that the computed marginal is a strong marginal [8, 20, 23],
but the set is not necessarily minimal.
As is the case for the h domain get marginal function, the means and variances corresponding to zero probability components are arbitrary numbers.
2

In the current implementation, all nodes must belong to the same junction tree.

126

The table returned by h node get distribution is owned by the application,

and it is the responsibility of the application to deallocate it (using h table
delete(73) ) after use.
Example 9.6 The following code prints out the components that form the (mixture) distribution for a continuous node. Each component is a (one-dimensional)
Gaussian distribution.
h_node_t n;
...
printf ("Distribution for %s:\n", h_node_get_name (n));
{
h_table_t t = h_node_get_distribution (n);
size_t k, s = h_table_get_size (t);
h_number_t *p = h_table_get_data (t);
for (k = 0; k < s; k++)
if (p[k] > (h_number_t) 0)
printf ("%g * Normal (%g, %g)\n",
(double) p[k],
(double) h_table_get_mean (t, k, n),
(double) h_table_get_variance (t, k, n));
(void) h_table_delete (t);
}
Note that we ignore the zero components of the distribution. (The table functions
used are described in Chapter 5.)

9.6

Retrieving expected utilities

In LIMIDs, we will want to retrieve the expected utilities associated with the
states of a node (usually a decision node). We might also be interested in
the overall expected utility of a decision problem.
First, evidence must be entered and propagated. Then, the functions below
can be used to retrieve expected utilities.
x

h number t h node get expected utility (h node t node, size t state)

If node is a discrete node, the expected utility associated with state is returned. If node is a utility node, the contribution of node to the overall
expected utility is returned (in this case, specify 0 as the value for state).

h number t h domain get expected utility (h domain t domain)

The overall expected utility of the decision problem represented by domain
is returned.
127

For both functions, a negative value is returned if an error occurs. But this
is of little use for error detection, since any real value is (in general) a valid
utility. Thus, errors must be checked for using the h error code(18) function.

9.7

Computing function values

The results of inference can be used as input to the functions associated with
real-valued function nodes.
x

h double t h node get value (h node t node)

Evaluate the function associated with node (which must be a real-valued
function node), and return the result.
If an error occurs, a negative number is returned. However, since a negative
value can be a valid result, error conditions must, in general, be checked for
using h error code(18) and related functions.
In order to successfully evaluate the function, the following conditions must
be satisfied:
if node belongs to a class, then it must not be an input node or a temporal clone;3
node must have a model;
all model nodes must be instantiated (as determined by the results of
inference);
the expression identified by the configuration of the model nodes must
be valid:
the expression must be non-NULL,
all subexpressions must have correct types (i.e., the types of the
operands must match the types expected by the operator), and
interval nodes must not appear in the expression;
there must be no computation failures (e.g., division by zero).
If non-real-valued function nodes appear in the expression, their values are
determined from the results of inference. In this case, the underlying domain must be compiled, and, moreover, the equilibrium must be sum, and
the evidence incorporation mode must be normal (see Section 10.1). These
conditions must also be satisfied, if the model has a non-empty set of model
nodes. (The conditions are not required in all other cases.)
3
The reason is that such nodes act as placeholders for nodes that wont be known until
the runtime domain has been constructed.

128

The values of non-real-valued function nodes are determined as follows:

the value of a labeled node is the label associated with the state to
which the node is instantiated (it is an error, if the node is not instantiated);
the value of a boolean node is the truth value associated with the
state to which the node is instantiated (it is an error, if the node is not
instantiated);
the value of a numbered or a CG node is the mean value of the node;
the value of a utility node is the expected utility of the node.
If the expression refers to the values of other (real-valued) function nodes,
then they must also be evaluated. This is done (recursively) according to
the procedure described above.
It is also possible to evaluate real-valued function nodes using the results of
simulation as input see h node get sampled value(145) .

9.8

Examining the evidence

The HUGIN API provides functions to access the evidence currently entered
to the nodes of a domain. Functions to determine the type of evidence (nonvacuous or likelihood) are also provided.
The node argument of the functions described below must be a discrete or a
continuous node. The functions having propagated in their names require
the underlying domain to be compiled. The functions having entered in
their names do not.
x

h number t h node get entered finding (h node t node, size t state)

Retrieve the finding value currently registered at the discrete node node for
state state. If an error is detected, a negative value is returned.

h number t h node get propagated finding

(h node t node, size t state)
Retrieve the finding value incorporated within the current junction tree potentials for state state of the discrete node node. If an error is detected, a
negative value is returned.

h double t h node get entered value (h node t node)

Retrieve the entered value for the continuous node node. If no value has
been entered, a usage error code is set and a negative number is returned.
129

However, since a negative number is a valid value for a continuous node,

checking for errors must be done using h error code(18) and friends.
h double t h node get propagated value (h node t node)

Retrieve the value that has been propagated for the continuous node node.
If no value has been propagated, a usage error code is set and a negative
number is returned. However, since a negative number is a valid value for
a continuous node, checking for errors must be done using h error code(18)
and friends.
h boolean t h node evidence is entered (h node t node)

Is the evidence potential, currently registered at node (which must be a

discrete or a continuous node), non-vacuous?
h boolean t h node likelihood is entered (h node t node)

Is the evidence potential, currently registered at node (which must be a

discrete or a continuous node), a likelihood?
Note: If node is a continuous node, false (zero) is returned.
h boolean t h node evidence is propagated (h node t node)

Is the evidence potential for node (which must be a discrete or a continuous node), incorporated within the current junction tree potentials, nonvacuous?
h boolean t h node likelihood is propagated (h node t node)

Is the evidence potential for node (which must be a discrete or a continuous node), incorporated within the current junction tree potentials, a likelihood?
Note: If node is a continuous node, false (zero) is returned.

9.9

Case files

When evidence has been entered to a set of nodes, it can be saved to a file.
Such a file is known as a case file. The HUGIN API provides functions for
reading and writing case files.
A case file is a text file. The format (i.e., syntax) of a case file can be described by the following grammar.
hCase filei

hNode findingi*

hNode findingi hNode namei:hValuei

130

 hState indexi | hLikelihoodi

hValuei

| hLabeli | hReal numberi | true | false

hState indexi

#hIntegeri

hLikelihoodi

( hNonnegative real numberi* )

where:
hState indexi is a valid specification for all discrete nodes. The state
index is interpreted as if specified as the last argument to h node select
state(121) for the named node.
hLikelihoodi is also a valid specification for all discrete nodes. A nonnegative real number must be specified for each state of the named
node (and at least one of the numbers must be positive).
hReal numberi is a valid specification for CG, numbered, and interval
nodes. For numbered and interval nodes, the acceptable values are
defined by the state values of the named node.
hLabeli is a valid specification for labeled nodes. The label (a doublequoted string) must match a unique state label of the named node.
If the contents of the string conform to the definition of a name (see
Section 13.8), the quotes can be omitted.
true and false are valid specifications for boolean nodes.
Comments can be included in the file. Comments are specified using the %
character and extends to the end of the line. Comments are ignored by the
case file parser.
Example 9.7 The following case file demonstrates the different ways to specify
evidence: A, B, and C are labeled nodes with states yes and no; D is a boolean
node; E is a numbered node; F is an interval node; and G is a CG node.
A:
B:
C:
D:
E:
F:
G:

"yes"
#1
(.3 1.2)
true
2
3.5
-1.4

% equivalent to "no"
% likelihood

Because yes is a valid name, the finding for A can instead be specified simply as:
A: yes

131

h status t h domain save case

(h domain t domain, h string t file name)
Create a case file named file name. (Note: If a file named file name already
exists and is not write-protected, it is overwritten.) The case file will contain
the evidence currently entered in domain. The contents is text conforming
to the above described format.
Note that if (some of) the nodes with evidence have not been assigned
names, then names will automatically be assigned (through calls to the h
node get name(42) function).

h status t h domain parse case

(h domain t domain, h string t file name,
void (error handler) (h location t, h string t, void ),
void data)
This function parses the case stored in the file with name file name. The
evidence stored in the case is entered into domain. All existing evidence in
domain is retracted before entering the new evidence.
The error handler and data arguments are used for error handling. This is
similar to the error handling done by the other parse functions. See Section 13.9 for further information.
In case of errors, no evidence will be entered.

132

Chapter 10

Inference
When evidence has been entered and the domain has been compiled, we
want to compute revised beliefs for the nodes of the domain. This process
is called inference. In HUGIN, this is done by a two-pass propagation operation on the junction tree(s). The two passes are known as CollectEvidence
and DistributeEvidence, respectively. The CollectEvidence operation proceeds
inwards from the leaves of the junction tree to a root clique, which has been
selected in advance. The DistributeEvidence operation proceeds outwards
from the root to the leaves.
This inference scheme is described in many places. See, for example, [8, 10,
13, 14, 15, 20].

10.1

Propagation methods

The collect/distribute propagation scheme can be used to compute many

different kinds of information.

10.1.1

Summation and maximization

One can think of a propagation as the computation of certain marginals of

the full joint probability distribution over all variables. As is well-known, the
distribution of an individual variable can be found by summing/integrating
out all other variables of this joint probability distribution.
However, we might also be interested in the probability, for each state of
a given variable, of the most probable configuration of all other variables.
Again, we can compute these probabilities from the joint probability distribution over all variables. But this time, we max out the other variables
(i.e., we take the maximum value over the set of relevant configurations).
133

It turns out that both kinds of marginals can be computed by the collect/distribute propagation method by a simple parametrization of the marginalization method.
When a propagation has been successfully completed, we have a situation
where the potentials on the cliques and the separators of the junction tree
are consistent, meaning that the marginal on a set S of variables can be
computed from the potential of any clique or separator containing S. We
also say that we have established equilibrium on the junction tree. The
equilibria, discussed above, are called sum-equilibrium and max-equilibrium,
respectively.
The HUGIN API introduces an enumeration type to represent the equilibrium. This type is called h equilibrium t. The values of this type are denoted by h equilibrium sum and h equilibrium max.

10.1.2

Evidence incorporation mode

The traditional way to incorporate (discrete) evidence into a junction tree is

to first multiply each evidence potential onto the potential of some clique;
when this has been done, the actual propagation is performed. This mode
of evidence incorporation is called the normal mode.
An alternative way to incorporate evidence is to multiply the evidence potentials onto the clique potentials during the propagation. If this is done in
the correct places, the equilibrium achieved will have the following property.
Assuming a sum-propagation, the resulting potential on a set V of variables
(clique, separator, or a single variable) will be the marginal probability for V
given evidence on all variables except the variables in V itself. Since this is
similar to the retraction of evidence (and accompanying propagation) for
each variable, this mode is known as the fast-retraction mode of evidence
incorporation.
A fast-retraction propagation can be useful to identify suspicious findings.
If the observation made on a variable has a very small probability in the
probability distribution obtained by incorporation of evidence on the other
variables, then quite likely something is wrong with the observation. (Another way to identify suspicious findings is to use the notion of conflict; see
Section 10.4).
If each item of evidence is a single-state observation of a single variable,
then the equilibrium achieved with a normal mode propagation will give
no useful information about the observed variables. In such cases, it would
be tempting to always choose a fast-retraction propagation. However, one
should be aware of the following facts: (1) a fast-retraction propagation
may fail due to logical relations in the domain model; (2) fast-retraction
propagations are not available for compressed domains, domains with continuous variables, or LIMIDs.
134

The fast-retraction propagation method is described in [7].

The HUGIN API introduces an enumeration type to represent the evidence
incorporation mode. This type is called h evidence mode t. The values of
this type are denoted by h mode normal and h mode fast retraction.

10.2

Propagation

Using the HUGIN API for inference involves several steps:

(1) Specify the evidence.
(2) Perform inference (i.e., update the junction tree potentials using the
specified evidence, and compute derived results such as beliefs and
expected utilities for all relevant nodes).
(3) Retrieve the results of inference.
The first and third steps are described in Chapter 9. The second step is
performed using the following function.
x

h status t h domain propagate

(h domain t domain, h equilibrium t equilibrium,
h evidence mode t evidence mode)
Establish the specified equilibrium on all junction trees of domain using evidence mode for incorporation of evidence. Revised beliefs are computed for
all (discrete and continuous) nodes. And if domain is a LIMID, revised expected utilities are also computed for all (discrete, continuous, and utility)
nodes.
If domain is a LIMID, then the specified evidence must form a valid evidence
scenario see Section 9.1.3.
If domain is compressed, evidence mode must be h mode normal. Unless all
nodes of domain are discrete chance nodes, equilibrium must be h equilibrium sum and evidence mode must be h mode normal.
DBN: If domain is a DBN runtime domain that has been triangulated using
h domain triangulate dbn(65) , then equilibrium must be h equilibrium sum.
The propagation operation requires a compiled domain see Chapter 7.
If models are used to specify tables, and the tables are not up-to-date with
respect to the models, then the tables are regenerated (using h node generate table(95) ).
If some tables (either because of direct modification or because of regeneration from a model) have changed since the most recent propagation operation, the equilibrium is computed (from scratch) using the updated tables.
In this case, conditional probability and policy tables are normalized (so that
135

the conditional probabilities corresponding to a given parent state configuration sum to 1). Note that the initial propagation (i.e., the one performed
by the compilation operation) considers all tables to be new.
Inference is performed separately for each junction tree. However, if there
is a functional trail from a node in one junction tree T1 to a node in another
junction tree T2 , then inference in T1 is performed before inference in T2 .
Since the purpose of such trails is usually to let the probability potentials
of some nodes in T1 depend on the values of some nodes in T2 , tables must
usually be regenerated between the individual inference operations.
In general, h domain propagate does not perform any unnecessary work. For
example, if the current evidence is a superset of the evidence propagated in
the most recent propagation (and the equilibrium and evidence incorporation mode of that propagation was equal to equilibrium and evidence mode,
and the set of node tables is unchanged), then an incremental propagation
is performed using d-separation analysis (see Section 9.4) to reduce the
extent of the DistributeEvidence pass.
However, if some currently propagated evidence has been retracted (and the
current evidence incorporation mode is normal), then the new equilibrium
must be computed either from scratch using the node tables or using a memory backup of the junction tree tables (see h domain save to memory(142) ).
Using a memory backup can significantly speed up the propagation in this
case.
If an error is detected during the propagation, the initial (that is, with no
evidence incorporated) distribution is established. This might fail, and if it
does, subsequent requests for beliefs and other quantities computed by a
propagation will fail.
The set of evidence is never changed by any of the propagation functions.
If the propagation fails, the error code (in addition to the general error
conditions such as usage and out-of-memory) can be:
h error inconsistency or underflow Some probability potential has degenerated into a zero-potential (i.e., with all values equal to zero). This
is almost always due to evidence that is considered impossible (i.e.,
its probability is zero) by the domain model. (In theory, it can also
be caused by underflow in the propagation process, but this rarely
happens in practice.)
h error overflow Overflow (caused by operations the purpose of which was
to avoid underflow) has occurred in the propagation process. This is
an unlikely error (but it is not entirely impossible it can occur in
very large naive Bayes models, for example). The error never occurs
if only one item of evidence is propagated at a time.
136

h error fast retraction A fast-retraction propagation has failed due to logical relations within the domain model.
h error invalid evidence (This only applies to LIMIDs.) The evidence scenario is invalid see Section 9.1.3.
It is also possible to perform inference on individual junction trees:
x

h status t h jt propagate
(h junction tree t jt, h equilibrium t equilibrium,
h evidence mode t evidence mode)
The meanings (including constraints) of the arguments and return value are
similar to the meanings and constraints of the arguments and return value
of the h domain propagate(135) function.

10.3 Inference in LIMIDs: Computing optimal policies

The propagation operation described above computes beliefs and expected
utilities based on the tables specified as part of the domain model: the conditional probability tables (associated with the chance nodes and the discrete function nodes), the utility tables (associated with the utility nodes),
and the policy tables (associated with the uninstantiated decision nodes).
Notice that the propagation operation doesnt change the policy tables it
simply uses the policy tables in the computations.
However, the pre-specified policy tables might not be optimal or, in other
words, changing the policy tables might improve the overall expected utility
of the decision problem.
The HUGIN API provides a function to compute policy tables that maximize
the overall expected utility. This function uses the SPU (Single Policy Updating) algorithm [24] to compute the policies. This is an iterative algorithm
that updates one policy at a time and terminates when all policies have converged (i.e., more iterations change nothing). The algorithm usually finds
the globally optimal policies, but it is possible that the algorithm may get
stuck at a local maximum.
The parents specified for the decision nodes determine which observations
should be taken into account when decisions are made. Ideally, we would
specify all observations to be taken into account, but this may not be practical because the size of a policy table is exponential in the number of parents.
We therefore often dont specify less important observations as parents of
decision nodes (for example, old observations are typically less important
than new ones) in order to reduce the size of the policy tables.
Unless all relevant information has been specified as parents, then it can be
useful to recompute policies whenever new information becomes available.
137

This is because the computations take all existing observations (in addition
to future observations specified as parents of decision nodes) into account
when policies are computed.
x

h status t h domain update policies (h domain t domain)

This function updates (using the SPU algorithm) the policies of uninstantiated decisions in domain such that the overall expected utility of the decision
problem is maximized (but note that this maximum is not guaranteed to be
a global maximum it is possible that it is a local maximum). Evidence is
taken into account (so the updated policies are effectively conditioned on
the evidence).
The junction tree potentials must be up-to-date with respect to the evidence,
the node tables (i.e., the CPTs, the current policies, and the utility tables)
and their models (if any). The equilibrium must be sum, and the evidence
incorporation mode must be normal (see Section 10.1).
The function only performs updates yielding improvements of the overall
expected utility. This implies that a policy is only changed for parent configurations matching the observations.
As the final step, the function performs a propagation using the new policies: The beliefs of nodes that depend on uninstantiated decision nodes,
the expected utilities of all nodes as well as the overall expected utility are
updated.

10.4

Conflict of evidence

An alternative to a fast-retraction propagation to identify suspicious findings

is to use the concept of conflict of evidence.
Given n items of evidence, e1 , . . . , en , we define the conflict of this set of
findings as
P(e1 ) P(en )
P(e1 , . . . , en )
This quantity can be computed as part of the propagation process with virtually no overhead.
More details can be found in [12]. Note that in the paper, the definition of
conflict of evidence includes the logarithm of the above ratio.
The current implementation of the HUGIN API does not support calculation
of conflict of evidence in LIMIDs and when evidence has been specified for
CG nodes.
The steps required for computing conflict of evidence are similar to those
for performing inference:
138

(1) Specify the evidence.

(2) Make sure the inference engine is in the initial state (i.e., sum-equilibrium with no evidence incorporated). This can be accomplished using
h domain reset inference engine(142) .
(3) Perform a sum-propagation (Section 10.2). This computes the conflict
of evidence.
(4) Retrieve the conflict of evidence.
DBN: In a DBN runtime domain that has been triangulated using h domain
triangulate dbn(65) , this procedure computes conflict values that are conditioned on the evidence that has been moved out of the time window by
h domain move dbn window(66) . That is, all probabilities in the definition of
conflict of evidence given above are conditioned on that evidence.
The conflict of evidence is retrieved using the following function.
x

h double t h domain get conflict (h domain t domain)

Return the conflict of evidence for domain computed during the most recent
propagation. If no evidence has been propagated, 1 is returned. In case of
errors, a negative number is returned.
The conflict of evidence for a domain is the product of the conflict of evidence for the individual junction trees of the domain. The following function returns the conflict of evidence for a single junction tree.

h double t h jt get conflict (h junction tree t jt)

Return the conflict of evidence for junction tree jt computed during the most
recent propagation. If no evidence has been propagated, 1 is returned. In
case of errors, a negative number is returned.
Example 10.1 The following code outlines the proper way of computing conflicts.
h_domain_t d;
...
/* enter evidence */
h_domain_reset_inference_engine (d);
h_domain_propagate (d, h_equilibrium_sum, h_mode_normal);
printf ("Conflict of evidence: %g\n",
h_domain_get_conflict (d));
This code first enters evidence into the domain, then ensures that the inference engine is in the initial state. Next, a propagation is performed. After the propagation,
the conflict of evidence is retrieved and printed.

139

10.5

The normalization constant

The final step of the collection phase of a propagation operation is to normalize the root clique potential.1 For a sum-propagation, the normalization
constant of this potential is the probability of the propagated evidence E:
= P(E)
For a max-propagation, the normalization constant is the probability of the
most probable configuration xE consistent with the propagated evidence E:
= P(xE )
This information is useful in many applications, so the HUGIN API provides
functions to access the normalization constant and its logarithm:
x

h double t h domain get normalization constant

(h domain t domain)

h double t h domain get log normalization constant

(h domain t domain)
If an error occurs, then h domain get normalization constant returns a negative number, and h domain get log normalization constant returns a positive
number.
The values returned by these functions are the normalization and the lognormalization constants for all of the propagated evidence. If the evidence
consists of several items of evidence, it makes no difference whether all
items are propagated together, or whether each item is entered and propagated in an incremental fashion.
DBN: If the time window has been moved, then the normalization constant
is the conditional probability of the evidence specified for the nodes in the
time window given the evidence that has been moved out of the window.
If the evidence is impossible (that is, the evidence has zero probability),
the propagation operation fails. This implies that the true normalization
constant is always positive.2 However, if many findings are propagated, the
normalization constant can become very small. If the normalization constant becomes smaller than the smallest positive floating-point number representable within the h double t type, it underflows and h domain get normalization constant returns 0. In this case, h domain get log normalization
1

In order to avoid underflow, local normalizations are performed in the separators as part
of CollectEvidence. The normalization constant also includes the constants used in the local
normalizations.
2
After a failed propagation operation, h domain get normalization constant returns 1.

140

constant can be used (this function returns the correct value for all successful propagations).3
If approximation is used, the normalization constant should be compared to
the error introduced by the approximation process see h domain get approximation constant(113) . If the probability of the evidence is smaller than
the approximation error, propagation within the original (unapproximated)
model should be considered (in order to get more accurate answers). See
also Section 7.7.
If likelihood evidence has been propagated, the normalization constant cannot, in general, be interpreted as a probability. As an example, consider
a binary variable: The likelihoods h 21 , 1i and h1, 2i yield the same beliefs,
but not the same normalization constant. However, see Example 9.1 and
Example 9.2 for cases where it makes sense to interpret the normalization
constant as a probability.
If CG evidence has been propagated, then the normalization constant is proportional to the density at the observed values of the continuous nodes (the
proportionality constant is the conditional probability of the discrete evidence given the CG evidence). The density depends directly on the scale
chosen for the continuous variables: Suppose that the scale of some continuous variable is changed from centimeter [cm] to millimeter [mm]. This
causes the density values for that variable to be reduced by a factor of 10.
Hence, the normalization constant should only be used to compare different
sets of findings.

10.6

Initializing the inference engine

As described in Section 10.2, the propagation operation does not, in general,

perform any unnecessary work: If possible, an incremental propagation is
performed. However, if evidence has been changed or retracted, then this
is not always possible. In this case, the inference engine must establish
junction tree potentials equivalent to the initial state of the inference engine
(i.e., with no evidence incorporated). This can be done by initializing the
inference engine from the node potentials (i.e., the conditional probability,
the policy, and the utility potentials).
As an alternative, a copy (often referred to as a memory backup) of the
initial state of the inference engine can be created. This memory backup
is then subsequently used for initializing the inference engine for the cases
where the initial state is needed. Using a memory backup can significantly
speed up inference.
3
If the log-normalization constant cannot be correctly computed, the propagation operation fails. But the inference engine tries very hard to avoid that.

141

h status t h domain save to memory (h domain t domain)

Create a memory backup of the belief and junction tree potentials of domain.
(This approximately doubles the memory consumption of domain.) The
equilibrium of the junction tree potentials must be sum (see Section 10.1)
with no evidence incorporated.
The operation is not supported for LIMIDs (because, in a LIMID, the initial
distribution is generally not useful as a starting point for an incremental
propagation).
The memory backup is kept up-to-date with respect to the node potentials: If
a node potential changes (either because of direct modification or because of
regeneration from a model), the memory backup is automatically refreshed.
On the other hand, the memory backup is deleted as part of the uncompilation and compression operations. Thus, it is necessary to recreate the
memory backup after a compilation or a compression operation.
The h domain save to memory function is typically called immediately after
a compilation (assuming there is no evidence) or a reset-inference-engine
operation.

h status t h domain reset inference engine (h domain t domain)

Establish the initial state of the inference engine of domain: sum-equilibrium
with no evidence incorporated.
This operation does not change the evidence (if any) specified for the nodes
of domain, but the beliefs of the nodes computed by the operation do not
take this evidence into account (that is, the beliefs are not conditioned on
the evidence).

h status t h domain initialize (h domain t domain)

Retract all evidence specified for the nodes of domain and establish the initial state of the inference engine.
This operation is equivalent to a h domain retract findings(122) operation followed by either a h domain reset inference engine(142) operation or a sumpropagation.
Example 10.2 The following code loads a domain and then repeatedly performs
some experiment on the domain, using the application-defined function perform
experiment, until some condition is satisfied. The domain is initialized before each
experiment starts, so that each experiment is carried out with the inference engine
in the initial state.
h_domain_t d = h_kb_load_domain ("mydomain.hkb", NULL);
int done = 0;
while (!done)

142

{
h_domain_initialize (d);
done = perform_experiment (d);
}
...

10.7

Querying the state of the inference engine

The HUGIN API provides several functions that enables the application to
determine the exact state of the inference engine. The following queries can
be answered:
Which type of equilibrium is the junction tree(s) currently in?
Which evidence incorporation mode was used to obtain the equilibrium?
Has evidence been propagated?
Has any likelihood evidence been propagated?
Is there any unpropagated evidence?
DBN: The functions described below for testing propagated evidence only
consider evidence for nodes in the time window. Evidence that has been
moved out of the time window using h domain move dbn window(66) are
ignored!
x

h boolean t h domain equilibrium is

(h domain t domain, h equilibrium t equilibrium)
Test if the equilibrium of all junction trees of domain is equilibrium.

h boolean t h jt equilibrium is
(h junction tree t jt, h equilibrium t equilibrium)
Test if the equilibrium of junction tree jt is equilibrium.

h boolean t h domain evidence mode is

(h domain t domain, h evidence mode t mode)
Test if the equilibrium of all junction trees of domain could have been obtained through a propagation using mode as the evidence incorporation
mode.
Note that without evidence, there is no difference between the equilibria
produced via normal or fast-retraction propagations.
143

h boolean t h jt evidence mode is

(h junction tree t jt, h evidence mode t mode)
Test if the equilibrium of junction tree jt could have been obtained through
a propagation using mode as the evidence incorporation mode.

h boolean t h domain evidence is propagated (h domain t domain)

Test if evidence has been propagated for domain.

h boolean t h jt evidence is propagated (h junction tree t jt)

Test if evidence has been propagated for junction tree jt.

h boolean t h domain likelihood is propagated

(h domain t domain)
Test if likelihood evidence has been propagated for domain.

h boolean t h jt likelihood is propagated (h junction tree t jt)

Test if likelihood evidence has been propagated for junction tree jt.

h boolean t h domain cg evidence is propagated

(h domain t domain)
Test if CG evidence has been propagated for domain.

h boolean t h jt cg evidence is propagated (h junction tree t jt)

Test if CG evidence has been propagated for junction tree jt.

h boolean t h domain evidence to propagate (h domain t domain)

Test if there is any node with changed evidence compared to the most recent
propagation (if any). If there was no previous propagation, this is equivalent
to testing if there is any node in domain with evidence.

h boolean t h jt evidence to propagate (h junction tree t jt)

Test if there is any node in the junction tree jt with new evidence as compared to the evidence propagated.

h boolean t h node evidence to propagate (h node t node)

Is the entered and propagated evidence for node (which must be a discrete
or a continuous node) different?

h boolean t h domain tables to propagate (h domain t domain)

Are there any nodes in domain having a (conditional probability, policy, or
utility) table that has changed since the most recent compilation or propagation operation?
144

h boolean t h jt tables to propagate (h junction tree t jt)

Similar to h domain tables to propagate, but specific to the junction tree jt.

10.8

Simulation

Often, we are interested in generating (sampling) configurations (i.e., vectors of values over the set of variables in the network) with respect to the
conditional distribution given the evidence.
h status t h domain simulate (h domain t domain)

If domain is compiled, sample a configuration with respect to the current

distribution on the junction tree(s). This distribution must be in sum-equilibrium with evidence incorporated in normal mode. Only propagated evidence is taken into account: Unpropagated evidence, and models and tables
that have changed since the most recent propagation operation are ignored.
If domain is not compiled, sample a configuration with respect to the distribution determined by the tables associated with the nodes of domain.
All uninstantiated chance, decision, and discrete function nodes must have
valid tables (see Section 2.6), and the set of nodes with evidence must form
an ancestral set of instantiated nodes (i.e., no likelihood or multi-state evidence, and if a chance node is instantiated, so are all of its discrete and
continuous parents). Tables that are not up-to-date with respect to their
models (see Chapter 6) are not regenerated.
Simulation is not supported for networks containing functional trails between
nodes that are not real-valued function nodes.
The sampled configuration is obtained using the following functions.4
h index t h node get sampled state (h node t node)

Retrieve the state index of node (which must be a discrete node) within the
configuration generated by the most recent call to h domain simulate. If an
error occurs, a negative number is returned.
h double t h node get sampled value (h node t node)

If node is a continuous node, then the value of node within the configuration
generated by the most recent call to h domain simulate is returned.
If node is a real-valued function node, the function associated with node is
evaluated using the configuration generated by h domain simulate as input
4

In order to avoid returning invalid values, simulation results are automatically invalidated when an uncompile operation is performed. It is an error to request values derived
from invalid simulation results.

145

(that is, if the function refers to a parent in an expression, then the sampled
value of that parent is used in the evaluation), and the result is returned.
If an error occurs, a negative number is returned. However, since negative
numbers can be valid in both cases, error conditions must be checked for
using h error code(18) and related functions.
x

h number t h node get sampled utility (h node t node)

Return the utility value in the potential table of node (which must be a utility
node) identified by the configuration generated by the most recent call to
h domain simulate. If an error occurs, a negative number is returned. But
since negative numbers are valid utility values, errors must be checked for
using h error code(18) and related functions.
The configurations generated by h domain simulate are not really random.
They are generated using a pseudorandom number generator producing a
sequence of numbers that although it appears random is actually completely
deterministic. To change the starting point for the generator, use the following function.

h status t h domain seed random

(h domain t domain, unsigned int seed)
Seed the pseudorandom number generator used by h domain simulate with
seed.
The pseudorandom number generator implemented in Hugin can also be
used directly through the following functions.

h double t h domain get uniform deviate (h domain t domain)

Use the pseudorandom number generator for domain to sample a real number from the uniform distribution over the interval [0, 1).

h double t h domain get normal deviate

(h domain t domain, h double t mean, h double t variance)
Use the pseudorandom number generator for domain to sample a real number from the normal (also known as the Gaussian) distribution with mean
mean and variance variance.

10.9

Value of information analysis

Consider the situation where a decision maker has to make a decision based
on the probability distribution of a hypothesis variable. It could, for instance, be a physician deciding on a treatment of a patient given the probability distribution of a disease variable. For instance, if the probability of
146

the patient suffering from the disease is above a certain threshold, then the
patient should be treated immediately. Prior to deciding on a treatment, the
physician may have the option to gather additional information about the
patient such as performing a test or asking a certain question. Given a range
of options, which option should the physician choose next? That is, which
of the given options will produce the most information? These questions
can be answered by a value of information analysis.
Given a Bayesian network model and a hypothesis variable, the task is to
identify the variable which is most informative with respect to the hypothesis variable.
Entropy and Mutual Information
The main reason for acquiring additional information is to reduce the uncertainty about the hypothesis under consideration. The selection of the variable to observe next (for example, the question to ask next) can be based
on the notion of entropy. Entropy is a measure of how much the probability
mass is scattered around on the states of a variable (the degree of chaos in
the distribution of the variable). In other words, entropy is a measure of
randomness. The more random a variable is, the higher its entropy will be.
The entropy H(X) of a discrete random variable X is defined as follows:
X
H(X) =
p(x) log p(x)
x

The maximum entropy log n (assuming X has n states) is achieved when

the probability distribution of X is uniform, while the minimum entropy 0
is achieved when all the probability mass is located on a single state. Thus,
0 H(X) log n.
Since entropy can be used as a measure of the uncertainty in the distribution
of a variable, we can determine how the entropy of a variable changes as
observations are made. In particular, we can identify the most informative
observation.
If Y is a random variable, then the conditional entropy H(X|Y) of X given Y
is computed as follows:
X
H(X|Y) =
p(y)H(X|Y = y) = H(X) I(X; Y)
y

where
I(X; Y) =

p(x, y) log

x,y

p(x, y)
p(x)p(y)

is the mutual information (also known as the cross entropy) of X and Y.

The conditional entropy H(X|Y) is a measure of the uncertainty of X given
147

an observation of Y, while the mutual information I(X; Y) is a measure of

the information shared by X and Y (i.e., the reduction in the entropy of X
obtained by observing Y). If X is the variable of interest, then I(X; Y) is a
measure of the value of observing Y. It follows that the variable Y to observe
next should be chosen such that I(X; Y) is maximized.
The HUGIN API provides functions for computing entropy and mutual information for discrete nodes. These functions use the natural logarithm as
the log function in the computations. The computations are done with
respect to the (joint) distribution computed by the most recent propagation
operation (implying that the underlying domain must be compiled). This
distribution must have been obtained by a sum-propagation using normal
evidence incorporation mode.
x

h double t h node get entropy (h node t node)

Compute the entropy of node (which must be a discrete node).

h double t h node get mutual information

(h node t node, h node t other)
Compute the mutual information of node and other (which must be discrete
nodes). The nodes must belong to the same domain, and this domain must
be compiled but not compressed.
If the nodes belong to distinct junction trees, then the network must not contain functional trails between nodes that are not real-valued function nodes.

10.10

Sensitivity analysis

Often there are one or more nodes (and associated beliefs) in a belief network that are regarded as outputs of the model. For example, the probability (belief) of a disease in a medical domain model.
In order to improve the robustness of the belief network model, it should
be investigated how sensitive the output probabilities are to variations in the
numerical parameters of the model (because these numerical parameters
are often imprecisely specified). The most influential parameters should be
identified, and effort should be directed towards reducing the impreciseness
of those parameters.
The process of identifying the most influential parameters of a belief network model and analyzing their effects on the output probabilities of the
model is known as sensitivity analysis [5, 6].
Let A be a (discrete chance) node in a belief network model being subjected
to sensitivity analysis, and let a be a state of A. For given evidence E, the
probability P(A = a|E) can be considered as a function of the conditional
148

probabilities in the CPTs of the chance nodes in the network. If the network
is a LIMID, P(A = a|E) can also be considered as a function of the conditional probabilities in the policies of the decision nodes.
Let B be a (discrete) node in the belief network, and let x be a conditional
probability parameter associated with B we shall refer to x as an input
parameter. This parameter corresponds to some state b of B and some state
configuration of the parents of B.
We wish to express P(A = a|E) as a function of x. When x varies, the other
conditional probabilities associated with B for parent configuration must
also vary (in order to satisfy the sum-to-1 constraint). The most common
assumption is that the other conditional probabilities vary according to a
proportional scheme:

P(b 0 |) =

if b 0 = b

1x
0
1 b| b |

if b 0 6= b

Here, b| is the initial (user-specified) value of x (and similarly for the other
input parameters).
Under this assumption (and also assuming that there are no functional trails
from B to A), it can be shown that the probability of the evidence E is a linear function of x:
P(E)(x) = x +

(10.1)

Likewise, the joint probability of the evidence E and the event A = a is a

linear function of x:
P(A = a, E)(x) = x +

(10.2)

x +
x +

(10.3)

We then conclude:
P(A = a|E)(x) =

This function is known as a sensitivity function.

If we know the constants (i.e., , , , and ) of this function, we can determine whether our estimate for x is precise enough: For example, we might
know that x belongs to some interval, and if the variations of P(A = a|E)(x)
for x ranging over that interval are acceptable, then we accept the estimate.
It turns out that the h, , , i-constants can be computed for all input parameters using only two propagations. The most efficient way to organize
the computations is therefore to provide two functions: One function to perform the propagations and to store the results of these propagations, and
another function to request the h, , , i-constants of a specific sensitivity
function (i.e., corresponding to a specific input parameter).
149

h status t h node compute sensitivity data

(h node t node, size t state)
This function computes the h, , , i-constants (or rather, compact data
from which these constants can be efficiently computed) of all sensitivity
functions (i.e., for all input parameters x) of the output probability
P(node = state|E)(x),
where node is a discrete chance node in a compiled domain, and E is the
current evidence. No evidence must be specified for node.
The junction tree potentials must be up-to-date with respect to the evidence,
the node tables (i.e., the CPTs, the policies, and the utility tables) and their
models (if any). The equilibrium must be sum, and the evidence incorporation mode must be normal (see Section 10.1).
Moreover, the network must not contain functional trails between nodes
that are not real-valued function nodes.
After calling the above function for producing sensitivity data, the following
function is called for each input parameter to be analyzed.

h status t h node get sensitivity constants

(h node t node, size t input, h number t , h number t ,
h number t , h number t )
This function retrieves the h, , , i-constants of the sensitivity function
corresponding to the specified input parameter: node must be a discrete
node, and input must be the index of an entry of the table of node. The
output probability of the sensitivity function was specified in the preceding
call to h node compute sensitivity data(150) .
The h, , , i-constants of the sensitivity function are returned in the h
number t arguments.
The user-specified value of the input parameter must not be 0 or 1 (because
the method used for computing the h, , , i-constants does not work in
these cases).
This function uses the sensitivity data produced by the preceding call to
h node compute sensitivity data. Certain operations on the underlying domain invalidate these data: An operation that uncompiles (Section 7.5)
the domain also invalidates the sensitivity data. Also, changing node tables
(i.e., the CPTs, the policies, and the utility tables) or their models (if any)
invalidate the sensitivity data.
If these conditions are not satisfied, the function returns an error.
Note that the h, , , i-constants returned by this function correctly define
the sensitivity function in (10.3), but the constants do not define the probabilities in (10.1) and (10.2) correctly (unless E is empty). One reason is that
150

the constants are scaled to avoid floating-point underflow: The scaling factor is P(E)1 (assuming the user-specified values for all input parameters).
Also, if = p and = p for some constant p, then the function might
return = = 0 and = p and = 1. This is the case if the specified input
parameter is not associated with a node in the sensitivity set (see below).
Example 10.3 This example uses the chest clinic network [25]. We shall determine the probability of lung cancer (L) given shortness of breath (D) as a function
of the prior probability of the patient being a smoker (S).
h_domain_t d = h_kb_load_domain ("asia.hkb", NULL);
h_number_t alpha, beta, gamma, delta;
h_node_t L, D, S;
... /* code to find the nodes */
h_domain_compile (d);
h_node_select_state (D, 0); /* set D to "yes" */
h_domain_propagate (d, h_equilibrium_sum, h_mode_normal);
h_node_compute_sensitivity_data (L, 0);
h_node_get_sensitivity_constants
(S, 0, &alpha, &beta, &gamma, &delta);
printf ("P(L=yes|D=yes)(x)"
" = (%g * x + %g) / (%g * x + %g)\n",
alpha, beta, gamma, delta);
This code produces the following output (reformatted to fit the width of the page):
P(L=yes|D=yes)(x) = (0.170654 * x + 0.0174324)
/ (0.535988 * x + 0.732006)
In this output, x refers to the prior probability of the patient being a smoker.

In many cases, only a small subset of the input parameters can influence the
output probability P(A = a|E). The set of nodes to which these input parameters are associated is known as the sensitivity set for A given E [6]. This set
of nodes can be identified using d-separation analyses (Section 9.4): Suppose that we add an extra parent X to node B. If X is d-connected to A
given E, then B belongs to the sensitivity set.
The h node compute sensitivity data(150) function identifies the sensitivity
set for the specified output probability given the available evidence. The
following function can be used to retrieve the sensitivity set.
x

h node t h domain get sensitivity set (h domain t domain)

Return the sensitivity set computed by the most recent call to h node compute sensitivity data, provided that the results of that call are still valid see
h node get sensitivity constants(150) for the list of conditions that invalidate
the sensitivity data.
The sensitivity set is returned as a list of nodes. This list is stored inside
domain, so the application must not attempt to deallocate it. Also note that
151

the list is invalidated by a subsequent call to h node compute sensitivity data

(i.e., it must be retrieved again, because it may have been reallocated).
Approximated domains (this typically includes compressed domains) are
treated specially. Because approximation can be seen as entering special
evidence in all cliques, we conservatively assume that all input parameters
may influence the specified output probability. In other words, the sensitivity set is taken to be the set of all nodes in the domain (regardless of what
the d-separation criterion tells us).
See Section 7.6 and Section 7.7 for more information regarding compression
and approximation.

Parameter tuning
In order to tune the behavior of a belief network, we often need to impose
constraints on the output probabilities of the network.
We recall the equation for the output probability P(A = a|E) as a function of
some input parameter x:
P(A = a|E)(x) =

x +
.
x +

Using this equation, we can find the values of x that satisfy a given constraint. In [5], the following types of constraints are considered:
(1) P(A = a|E) P(A 0 = a 0 |E)

(D IFFERENCE )

(2) P(A = a|E) / P(A 0 = a 0 |E)

(R ATIO)

(3) P(A = a|E)

(S IMPLE)

where P(A = a|E) and P(A 0 = a 0 |E) are output probabilities (referring to the
same variable or to two different variables). The relation can also be .
If x is associated with a node belonging to the sensitivity sets of both output
probabilities, the sensitivity functions (as computed by h node compute sensitivity data(150) or h domain compute sensitivity data(153) ) are:
P(A = a|E)(x) =

1 x + 1
x +

and

P(A 0 = a 0 |E)(x) =

2 x + 2
x +

For the D IFFERENCE constraint, we get (because x + is nonnegative):

P(A = a|E) P(A 0 = a 0 |E)

(1 2 )x 1 + 2 +

If 1 2 is positive, we get a lower bound for x. If 1 2 is

negative, we get an upper bound.
152

If x is not associated with a node belonging to the sensitivity sets of both output probabilities, then at least one of the probabilities is a constant. If both
are constants, the solution is trivial. If only one is a constant, the D IFFER ENCE constraint reduces to a S IMPLE constraint.
For the R ATIO constraint, we get:
P(A = a|E) / P(A 0 = a 0 |E)

(1 2 )x 1 + 2

Depending on the sign of 1 2 , we get either a lower or an upper bound.

Finally, for the S IMPLE constraint, we get:
P(A = a|E)

(1 )x 1 +

Again, we get either a lower or an upper bound.

This can be generalized to any number of constraints (by forming the intersection of solution intervals) and any type of solvable constraint. Solving
a set of parameter constraints is done in two stages.
The first stage consists of computing sensitivity data corresponding to all of
the output probabilities involved in the constraints. Let the output probabilities be P(A0 = a0 |E), P(A1 = a1 |E), . . . , P(An1 = an1 |E). These probabilities are encoded using two lists: a NULL-terminated list of nodes and
a list of state values, such that the state value at position i in the second
list is associated with the node at position i in the first list. A given output
probability is then later referred to by its position in these lists.
The second stage computes the solution to the constraints.
x

h status t h domain compute sensitivity data

(h domain t domain, h node t nodes, const size t states)
Compute sensitivity data for all of the specified output probabilities. All of
the nodes must belong to domain. The usage conditions of h node compute
sensitivity data(150) must be satisfied for all output probabilities.
Note: h node compute sensitivity data is implemented in terms of this function (implying that each function overwrites the sensitivity data computed
by the other function).

h status t h node get sensitivity constants by output

(h node t node, size t input, size t output, h number t ,
h number t , h number t , h number t )
This function retrieves the h, , , i-constants of the sensitivity function
corresponding to the specified input parameter and the specified output
probability: node must be a discrete node, input must be the index of an
entry of the table of node, and output must identify an output probability
specified in the preceding call to h domain compute sensitivity data(153) .
153

See h node get sensitivity constants(150) for the remaining usage conditions
and the semantics of this function.
Note: h node get sensitivity constants is implemented in terms of this function.
x

h node t h domain get sensitivity set by output

(h domain t domain, size t output)
Return the sensitivity set corresponding to the specified output probability:
output refers to one of the output probabilities specified in the most recent
call to h domain compute sensitivity data(153) .
See also h domain get sensitivity set(151) .
Note: h domain get sensitivity set is implemented in terms of this function.

10.11

Most probable configurations

Steffen Lauritzen has proposed a Monte Carlo algorithm for finding the most
probable configurations of a set of discrete nodes given evidence on some of
the remaining nodes.
Let A be the set of nodes for which we seek the most probable configurations. The goal is to identify all configurations with probability at least pmin .
The algorithm can be outlined as follows.
A sequence of configurations of A is sampled from the junction tree.
[Configurations of A are obtained by ignoring the sampled values of
the remaining nodes.] Repetitions are discarded.
The probability of each (unique) configuration in the sequence is computed. Let ptotal be the total probability of all (unique) configurations.
If 1 ptotal < pmin , the algorithm terminates.
The most probable configuration is also known as the maximum a posteriori
(MAP) configuration. The above algorithm thus solves a more general form
of the MAP configuration problem.
The basic assumption of the algorithm is that most of the probability mass
is represented by a small set of configurations. If this is not the case, the
algorithm could run for a long time.
If the size of A is small, it is probably more efficient to compute the joint
probability table of A using h domain get marginal(126) . From this table, it
is easy to identify the most probable configurations.
x

h status t h domain find map configurations

(h domain t domain, h node t nodes, h double t pmin )
Find all configurations of nodes with probability at least pmin (0 < pmin 1).
The nodes list must be a nonempty list of distinct discrete nodes belonging to
154

domain. The domain must be compiled, and the distribution on the junction
tree(s) must be in sum-equilibrium with evidence incorporated in normal
mode (see Section 10.1).
The current implementation imposes the following additional restrictions.5
LIMIDs are not supported (that is, domain must not contain decisions).
Evidence must not be specified for any of the nodes in nodes.
The junction tree potentials must be up-to-date with respect to the
evidence, the CPTs and their models (if any).
The network must not contain functional trails between nodes that are
not real-valued function nodes.
The probability pmin should be reasonable (like 0.01 or higher). Otherwise,
the algorithm could run for a long time.
The results of a call to h domain find map configurations are retrieved using
the functions described below. The results remain available until domain is
uncompiled.
h count t h domain get number of map configurations
(h domain t domain)

This function returns the number of configurations found by the most recent
successful call to h domain find map configurations (with domain as argument). If no such call has been made (or the results of the call are no longer
available), a negative number is returned.
Let n be the number of configurations with probability at least pmin as
specified in the call to h domain find map configurations. The configurations are identified by integer indexes 0, . . . , n 1, where index 0 identifies
the most probable configuration, index 1 identifies the second-most probable configuration, . . . , and index n 1 identifies the least probable of the n
configurations.
size t h domain get map configuration
(h domain t domain, size t index)

This function returns the configuration identified by index among the configurations with probability at least pmin as specified in the most recent
successful call to h domain find map configurations (as explained above). If
an error occurs, NULL is returned.
5

An alternative implementation is under consideration. This implementation might impose a different set of restrictions (e.g., requiring that domain not being compressed, but
removing the other additional restrictions).

155

A configuration is stored as a list of state indexes (an array of size t values):

The kth element of a configuration is a state index for the kth node in the
nodes list as specified in the call to h domain find map configurations.
The configuration is stored within the domain data structure, so the application must not deallocate or modify the configuration. The HUGIN API
deallocates the configurations when domain is uncompiled.
x

h double t h domain get probability of map configuration

(h domain t domain, size t index)
Return the probability of the configuration returned by h domain get map
configuration (domain, index). If an error occurs, a negative number is returned.

10.12

Explanation

The explanation facilities in the HUGIN API consist of computing the impact
of subsets of the given evidence on a single hypothesis [18, section 10.1.3]
or on one hypothesis versus an alternative hypothesis [18, section 10.1.4].
Impact of evidence subsets on a single hypothesis. We wish to investigate how
different subsets of the evidence support (or do not support) a given hypothesis. Let E be the set of evidence (a single piece of evidence e E consists
of evidence on a unique variable), and let X = x be the hypothesis (where X
is a discrete variable, and x is a state of X).
We compute the impact score of a subset E 0 of the evidence on the hypothesis X = x as the normalized likelihood of the hypothesis given the evidence E 0 :
P(E 0 |X = x)
P(X = x|E 0 )
=
0
P(E )
P(X = x)
We assume P(E 0 ) > 0 and P(X = x) > 0.
The normalized likelihood is a measure of the impact of the evidence on the
hypothesis. By comparing the normalized likelihoods of different subsets of
the evidence, we compare the impacts of the subsets on the hypothesis.
Discrimination of competing hypotheses. We wish to investigate how different
subsets of the evidence support (or do not support) different (competing)
hypotheses. Let E be the set of evidence (a single piece of evidence e E
consists of evidence on a unique variable), and let X = x and Y = y be the
competing hypotheses (where X and Y are discrete variables, and x and y
are states of X and Y, respectively).
We shall use the Bayes factor (or Bayesian likelihood ratio) B as the impact
score of a subset E 0 of the evidence on the (primary) hypothesis X = x versus
156

the (alternative) hypothesis Y = y:

B=

posterior odds ratio

P(X = x|E 0 )/P(Y = y|E 0 )
=
prior odds ratio
P(X = x)/P(Y = y)

We assume P(X = x) > 0, P(Y = y) > 0, and P(E 0 ) > 0.

We see that
B > 1 if the evidence E 0 provides more support for X = x than for Y = y.
B < 1 if the evidence E 0 provides less support for X = x than for Y = y.
B = 1 if the evidence E 0 does not provide any information for discriminating between X = x and Y = y.
x

h status t h node compute explanation data

(h node t X, size t x, h node t Y, size t y,
size t max subset size)
Compute explanation data for all nonempty evidence subsets of size max
subset size.6 The nodes X and Y must be discrete nodes in a compiled domain, and x and y must be valid states of X and Y, respectively. The node Y
can be NULL in that case, y must be 0.
The junction tree potentials must be up-to-date with respect to the evidence,
the node tables (i.e., the CPTs, the policies, and the utility tables) and their
models (if any). The equilibrium must be sum, and the evidence incorporation mode must be normal (see Section 10.1).
The probabilities of the hypotheses in the initial distribution (that is, before
any evidence is incorporated) must be positive.
LIMIDs are not supported (that is, the domain must not contain decisions).
If Y is non-NULL, Bayes factors for the hypothesis X = x versus the alternative
hypothesis Y = y are computed. If Y is NULL, then normalized likelihoods are
computed.
The explanation data is stored within the domain to which the nodes belong.
A successful call to the h node compute explanation data function overwrites
any existing explanation data (i.e., results produced by a previous call to this
function) stored within the domain.
Subsets of the evidence for which both hypotheses have zero probability are
excluded from the results. The empty subset is also excluded since it doesnt
provide any information.
6
The max subset size argument makes it possible to reduce the demand for computation
resources (both time and space) of this function when there is evidence on many nodes.

157

When the domain to which X and Y belong is uncompiled, the explanation

data stored in the domain is also deallocated.
x

h count t h domain get number of explanations

(h domain t domain)
This function returns the number of explanations found by the most recent
successful call to h node compute explanation data (with nodes in domain as
arguments). If no such call has been made (or the results of the call are no
longer available), a negative number is returned.
Let n be the number of explanations. The explanations are identified by
integer indexes 0, . . . , n 1, where index 0 identifies the explanation with
the highest score, index 1 identifies the explanation with the second-highest
score, . . . , and index n 1 identifies the explanation with the lowest score.

h node t h domain get explanation

(h domain t domain, size t index)
This function returns the explanation (i.e., a subset of the evidence nodes)
identified by index. If an error occurs, NULL is returned.
The function returns the explanation using a node list stored within domain,
which means that the application must not free the list. The list is allocated
by h node compute explanation data(157) and is deallocated together with
the rest of the explanation data. The list is overwritten by each (successful)
call to the h domain get explanation function.

h number t h domain get explanation score

(h domain t domain, size t index)
Return the impact score of the explanation that is returned by h domain get
explanation(158) with domain and index as arguments. If an error occurs, a
negative number is returned.

158

Chapter 11

Sequential Updating of
Conditional Probability Tables
This chapter describes the facilities for using data to sequentially update the
conditional probability tables for a domain when the graphical structure and
an initial specification of the conditional probability distributions have been
given in advance.
Sequential updating makes it possible to update and improve these conditional probability distributions as observations are made. This is especially
important if the model is incomplete, the modeled domain is drifting over
time, or the model quite simply does not reflect the modeled domain properly.
The sequential learning method implemented (also referred to as adaptation) was developed by Spiegelhalter and Lauritzen [35]. See also Cowell et
al [8] and Olesen et al [29].
Spiegelhalter and Lauritzen introduced the notion of experience. The experience is the quantitative memory which can be based both on quantitative
expert judgment and past cases. Dissemination of experience refers to the
process of calculating prior conditional distributions for the variables in the
belief network. Retrieval of experience refers to the process of calculating
updated distributions for the parameters that determine the conditional distributions for the variables in the belief network.

11.1

Experience counts and fading factors

The adaptation algorithm updates the conditional probability distributions

of a belief network in the light of inserted and propagated evidence. Adaptation can be applied to discrete chance variables only.
159

The experience for a given node is represented as a set of experience counts

0 , . . . , n1 one count for each configuration of the (discrete) parents
of the node: i represents the number of times the (discrete) parents have
been observed to be in the ith configuration. However, note that the counts
dont have to be integers (non-integer counts can arise because of incomplete case data).
The experience counts are stored in a table:
h table t h node get experience table (h node t node)

This function returns the experience table of node (which must be a chance
node).1 If node doesnt have an experience table, then one will be created.
The nodes of the experience table are the discrete parents of node, and the
order of the nodes in the experience table is the same as the order in the
conditional probability table of node.
OOBN: node must not be an output clone.
When an experience table is created, it is filled with zeros. Zero is an invalid
experience count for the adaptation algorithm, so positive values must be
stored in the table before adaptation can take place. If a database of cases
is available, the EM algorithm can be used to get initial experience counts.
The adaptation algorithm only adapts conditional distributions corresponding to parent configurations with positive experience counts. All other configurations (including all configurations for nodes with no experience table)
are ignored. This convention can be used to turn on/off adaptation at the
level of individual parent configurations: Setting an experience count to a
positive number turns on adaptation for the associated parent configuration;
setting the experience count to zero or a negative number turns it off.
Note that the table returned by h node get experience table is the table stored
within node (and not a copy of that table). This implies that the experience
counts for node can be modified using functions that provide access to the
internal data structures of tables (see Chapter 5).
Experience tables can be deleted using the h table delete(73) function. Deleting an experience table turns off adaptation for the node associated with the
table.
h boolean t h node has experience table (h node t node)

This function tests whether node has an experience table.

The adaptation algorithm also provides an optional fading feature. This
feature reduces the influence of past (and possibly outdated) experience in
order to let the domain model adapt to changing environments. This is
1

Although adaptation is only possible for discrete chance nodes, experience tables are
also used to control the EM algorithm, and the EM algorithm applies to both discrete and
continuous chance nodes.

160

done by discounting the experience count i by a fading factor i , which is

a positive real number less than but typically close to 1. The true fading
amount is made proportional to the probability of the parent configuration
in question. To be precise: If the probability of the ith parent configuration
given the propagated evidence is pi , then i is multiplied by (1 pi ) + pi i
before adaptation takes place. Note that experience counts corresponding
to parent configurations that are inconsistent with the propagated evidence
(i.e., configurations with pi = 0) remain unchanged.
The fading factor i can be set to 1: this implies that cases are accumulated
(that is, no fading takes place). Setting i to a value greater than 1 or less
than or equal to 0 will disable adaptation for the ith parent configuration
(just as setting i to an invalid value will).
The fading factors are also stored in a table:
x

h table t h node get fading table (h node t node)

This function returns the fading table of node (which must be a discrete
chance node). If node doesnt have a fading table, then one will be created.
The order of the nodes in the fading table is the same as the order of the parents of node in the conditional probability table of node (which is identical
to the order of nodes in the experience table).
OOBN: node must not be an output clone.
When a fading table is created, all entries are initially set to 1. This has the
same affect as if no fading table were present. To get fading, values between
0 and 1 must be stored in the table.
The table returned by h node get fading table is the table stored within node
(and not a copy of that table). This implies that the fading factors for node
can be modified using functions that provide access to the internal data
structures of tables (see Chapter 5).
Like experience tables, fading tables can also be deleted using the h table
delete(73) function. Note that fading tables are not automatically deleted
when the corresponding experience tables are deleted. The fading tables
must be explicitly deleted.

h boolean t h node has fading table (h node t node)

This function tests whether node has a fading table.
Example 11.1 The following code loads the Asia domain [25], enables adaptation for all nodes except E (we delete the experience table of E, if it has one): If
some node (besides E) doesnt have an experience table, we create one and set all
entries of the table to 10.
h_domain_t d = h_kb_load_domain ("Asia.hkb", NULL);
h_node_t E = h_domain_get_node_by_name (d, "E");

161

h_node_t n = h_domain_get_first_node (d);

for (; n != 0; n = h_node_get_next (n))
if (n != E && !h_node_has_experience_table (n))
{
h_table_t t = h_node_get_experience_table (n);
h_number_t *data = h_table_get_data (t);
size_t k = h_table_get_size (t);
for (; k > 0; k--, data++)
*data = 10.0;
}
if (h_node_has_experience_table (E))
h_table_delete (h_node_get_experience_table (E));

11.2

Updating tables

When experience tables (and optionally fading tables) have been created
and their contents specified, then the model is ready for adaptation.
An adaptation step consists of entering evidence, propagating it, and, finally,
updating (adapting) the conditional probability and experience tables. The
last substep is performed using the following function.
x

h status t h domain adapt (h domain t domain)

This function updates (adapts), for all discrete chance nodes of domain, the
experience count (retrieval of experience) and the conditional probability
distribution (dissemination of experience) for all parent configurations having a valid experience count and a valid fading factor.
This adaptation is based on the propagated evidence (hence domain must be
compiled). The evidence must have been propagated with equilibrium equal
to sum and evidence-incorporation-mode equal to normal (Section 10.1).
Moreover, the junction tree potentials must be up-to-date with respect to the
node tables and their models (the h domain tables to propagate(144) function tests this condition). Note that the latter condition implies that the h
domain adapt function cannot be (successfully) called before the updated
distributions have been incorporated into the junction tree potentials (by
either a propagation or a reset-inference-engine operation).
Example 11.2 The following code
h_domain_t d = h_kb_load_domain ("Asia.hkb", NULL);
h_node_t n = h_domain_get_node_by_name (d, "S");

162

h_node_select_state (n, 0);

h_domain_propagate (d, h_equilibrium_sum, h_mode_normal);
h_domain_adapt (d);
...
loads the Asia domain [25], enters the observation that node S is in state 0 (yes),
propagates the evidence, and updates the experience and conditional probability
tables of the domain (using h domain adapt). We assume that suitable experience
(and possibly fading) tables have already been specified.

If the experience count for some parent configuration is (or can be expected
to be) very large (104 or more) or the fading factor is very close to 1 (1104
or closer), then it is recommended that a double-precision version of the
HUGIN API is used.

163

164

Chapter 12

Learning Network Structure

and Conditional Probability
Tables
Chapter 11 describes the facilities for adapting the conditional probability
distributions of a domain as new observations are made. This is known as
sequential learning.
However, the sequential learning method requires that a complete belief
network model including an initial assessment of the conditional probabilities is given. This chapter describes the HUGIN API facilities for using data
(a set of cases) to learn the network structure as well as the conditional
probability distributions of a belief network model. This is known as batch
learning. Batch learning requires that all data is available when the learning process starts, whereas sequential learning allows the knowledge to be
updated incrementally.
The method for learning the network structure is the PC algorithm, developed by Spirtes and Glymour [36]. A similar algorithm (the IC algorithm)
was independently developed by Verma and Pearl [32].
The method for learning the conditional probability distributions (a method
based on the EM algorithm) was developed by Lauritzen [21]. See also
Cowell et al [8].

12.1

Data

An assignment of values to some or all of the nodes of a domain is called

a case. If values have been assigned to all nodes, the case is said to be
complete; otherwise, it is said to be incomplete. The data used by the learning
procedure is comprised of a set of cases.
165

Note that the mechanism for entering cases described in this section is intended for case sets that fit in main memory. The learning algorithms currently provided by the HUGIN API assume that the data is stored in main
memory. Also note that case data is not saved as part of the HUGIN KB file
produced by h domain save as kb(49) .
The cases are numbered sequentially from 0 to N 1, where N is the total
number of cases. The first case gets the number 0, the second case gets the
number 1, etc.
x

h status t h domain set number of cases

(h domain t domain, size t count)
This function adjusts the storage capacity for cases of domain to count. Let
the current capacity be m. The contents of the cases numbered 0 up to
min(count, m) 1 are unchanged. If count > m, then the contents of the
cases numbered m to count 1 are set to unknown. If count < m, then the
cases numbered count to m 1 are deleted. In particular, setting count to 0
deletes all cases.
The following function is provided for convenience (e.g., for use when reading a file of cases where the number of cases is not known in advance).

h index t h domain new case (h domain t domain)

Allocate storage within domain to hold a new case. If successful, the function returns the index of the new case. If an error occurs, a negative number
is returned.

h count t h domain get number of cases (h domain t domain)

Returns the number of cases currently allocated for domain.

h status t h node set case state

(h node t node, size t case index, size t state index)
Specify the state value of the discrete node node associated with case case
index to be state index; state index must be an integer identifying a state of
node (similar to the last argument of the function h node select state(121) ).
If the number of states of node is subsequently decreased (such that state
index becomes invalid), then the learning algorithms will treat the data as
unknown/missing.
In order to reduce memory consumption, the value of state index must be
less than or equal to 32767 (that is, a signed 2-byte quantity is allocated for
each case state index).

h index t h node get case state (h node t node, size t case index)
Retrieve the state value of the discrete node node associated with case case
index. If an error occurs or no state value (or an invalid state value) has
been specified, a negative number is returned.
166

Case data for continuous nodes is specified using the following functions.
x

h status t h node set case value

(h node t node, size t case index, h double t value)
Set the value associated with the continuous node node in case case index to
value.

h double t h node get case value (h node t node, size t case index)
Retrieve the value of the continuous node node associated with case case
index. If an error occurs, a negative number is returned, but this cannot be
used for error detection, since any real value is a valid value. Instead, the
h error code(18) function must be used.
The next two functions apply to both discrete and continuous nodes.

h status t h node unset case (h node t node, size t case index)

Specify that the value of node in case case index is unknown.
Note that this function should rarely be needed, since the state values for all
nodes of a newly created case are unknown, and also the value of a newly
created node will be unknown in all cases.

h boolean t h node case is set (h node t node, size t case index)

Test whether the value of node in case case index is currently set.
In large data sets, some cases may appear more than once. Instead of entering the case each time it appears, a count may be associated with the case.
This count must be nonnegative (a zero case-count means that the case will
be ignored by the learning algorithms), but it doesnt have to be an integer.

h status t h domain set case count

(h domain t domain, size t case index, h number t case count)
Set the case-count for the case with index case index in domain to case count.

h number t h domain get case count

(h domain t domain, size t case index)
Retrieve the case-count associated case case index in domain.
If no case-count has been associated with a case, the count defaults to 1.
The case-counts have no influence on the value returned by h domain get
number of cases(166) .
The learning algorithms operate on the data set as a whole. But sometimes
it can be useful to perform inference (or some other task) using a specific
case in the data set. The following function assists with such tasks.
167

h status t h domain enter case

(h domain t domain, size t case index)
Enter the case data from case case index as evidence into domain. All existing evidence in domain is retracted before entering the case.

12.2

Data files

When a set of cases has been entered as described in the previous section,
it can be saved to a file. Such a file is known as a data file. The HUGIN API
provides functions for reading and writing data files.
A data file is a text file. The format (i.e., syntax) of a data file can be
described by the following grammar.
hData filei

hHeaderihCasei*

hHeaderi

#hSeparatorihNode listi | hNode listi

hSeparatori , | hEmptyi
hNode listi

hNode namei | hNode listihSeparatorihNode namei

hCasei

hCase countihSeparatorihData listi | hData listi

hCase counti hNonnegative real numberi

hData listi

hDatai | hData listihSeparatorihDatai

hDatai

hValuei | * | ? | hEmptyi

where:
The hHeaderi must occupy a single line in the file. Likewise, each
hCasei must occupy a single line.
If # is the first element of hHeaderi, then each hCasei must include a
hCase counti.
Each hCasei must contain a hDatai item for each node specified in the
hHeaderi. The ith hDatai item (if it is a hValuei) in the hData listi must
be valid (as explained in Section 9.9) for the ith node in the hNode listi
of the hHeaderi.
If hDatai is *, ?, or hEmptyi, then the data is taken as missing.
If hSeparatori is hEmptyi, then none of the separated items is allowed
to be hEmptyi.
hValuei is as defined in Section 9.9, except that the hLikelihoodi alternative is not allowed.
168

Comments can be included in the file. Comments are specified using the %
character and extends to the end of the line. Comments behave like newline
characters. Empty lines (after removal of blanks, tabs, and comments) are
ignored by the data file parser (i.e., they do not represent empty cases).
Example 12.1 Here is a small set of cases for the Asia domain [25].
#

1
1
1
1
2
1
1
1

"yes"
"no"
"no"
"no"
"yes"
"yes"
"yes"
"no"

"no"
"yes"
"yes"
"no"
"yes"
"no"
"yes"
"no"

"no"
"yes"
"yes"
"yes"
*
"no"
"yes"
"no"

"no"
"no"
"yes"
"yes"
"no"
*
"yes"
*

The first line lists the nodes, and the remaining lines each describe a case. The first
case corresponds to a non-smoking patient, that has been to Asia recently, does
not have shortness of breath, and the X-ray doesnt show anything. The last case
corresponds to a non-smoking patient, that has not (recently) been to Asia, does
not have shortness of breath, and the X-ray is not available. Similarly for the other
cases.
Note the extra (optional) initial column of numbers: These numbers are case
counts. The number 2 for the fifth case indicates that this case has been observed
twice; the other cases have only been observed once. The presence of case counts
is indicated by the # character in the header line.

Note the distinction between case files (Section 9.9) and data files: A case
file contains exactly one case, it may contain likelihood data, and reading a
case file means loading the case data as evidence. A data file, on the other
hand, can contain arbitrarily many cases, but likelihood data is not allowed,
and reading a data file (using h domain parse cases described below) loads
the case data using the facilities described in Section 12.1.
x

h status t h domain parse cases

(h domain t domain, h string t file name,
void (error handler) (h location t, h string t, void ),
void data)
This function parses the cases stored in the file with name file name. The
cases will be stored in domain using the facilities described in Section 12.1.
Existing cases in domain are not modified.
The error handler and data arguments are used for error handling. This is
similar to the error handling done by the other parse functions. See Section 13.9 for further information.
If an error occurs, no cases will be added to domain.
169

The h domain save cases function saves case data stored in a domain.
h status t h domain save cases
(h domain t domain, h string t file name, h node t nodes,
h index t cases, h boolean t include case counts,
h string t separator, h string t missing data)

Save (some of) the case data stored in domain as a data file named file name.
(Note: If a file named file name already exists and is not write-protected, it
is overwritten.) The format and contents of the file are controlled by several
arguments:
nodes is a non-empty NULL-terminated list of (distinct) nodes. Moreover,
all nodes must be discrete or continuous nodes belonging to domain.
The list specifies which nodes (and their order) that are saved.
Note: If (some of) the nodes do not have names, they will be assigned
names (through calls to the h node get name(42) function).
cases is a list of case indexes (which must all be valid), terminated by 1.
The list specifies which cases (and their order in the file) that must
be included. Duplicates are allowed (the case will be output for each
occurrence of its index in the list).
When a case is output, the associated case count is output unmodified:
If the case has case count n, then it is also n in the generated file (not
n/2 or something like that).
can be passed for cases. This will cause all cases to be output (in
the same order as stored in domain).
NULL

include case counts is a boolean controlling the presence of case counts in

the generated file:
If it is true, case counts will be included (even if they are all 1).
If it is false, they are not included. This is only allowed if all
the selected cases have integer-valued case counts because, instead of writing the case count to the file, the case is repeated
as many times as specified by the case count. Note: If the case
count is zero, the case will be omitted from the file.
separator is a string that is output between node names in the header and
between data items (and after the case count) in a case. If the generated file is to be read by h domain parse cases, then separator must
be non-empty, must contain at most one comma, and the remaining
characters must be blanks or tabs.1
1

If the data file is to be read by other applications, it can be useful to use a different
separator and/or a different missing data indicator. Therefore, these restrictions are not
enforced.

170

missing data is a string that is output if no data is specified for a given

node in a given case. If the generated file is to be read by h domain
parse cases, then the following restrictions apply: missing data must
contain at most one of the * or ? characters; if missing data does not
contain one of these characters, then separator must contain a comma;
the remaining characters must be blanks or tabs.1
Example 12.2 Let Asia.dat be a data file with contents as shown in Example 12.1. The following code loads the Asia domain [25], parses the Asia.dat
data file, and generates a new data file (New.dat) containing a subset of the data.
[See Example 13.25 for an appropriate definition of the error handler used by the
parse function.]
h_domain_t d = h_kb_load_domain ("Asia.hkb", NULL);
h_index_t cases[5] = { 0, 2, 4, 6, -1 };
h_node_t nodes[4];
nodes[0]
nodes[1]
nodes[2]
nodes[3]

=
=
=
=

h_domain_get_node_by_name (d, "S");

h_domain_get_node_by_name (d, "D");
h_domain_get_node_by_name (d, "X");
NULL;

h_domain_parse_cases
(d, "Asia.dat", error_handler, "Asia.dat");
h_domain_save_cases
(d, "New.dat", nodes, cases, 0, ",\t", "");
When this code is executed, a new data file (New.dat) is generated. It has the
following contents:
S,

D,

"no",
"yes",
"yes",
"yes",
"yes",

"no",
"yes",
,
,
"yes",

"no"
"yes"
"no"
"no"
"yes"

Note that the case with index 4 (the fifth case) from the input data file is repeated
twice in the output data file. This is because that case has case count 2 in the input
data.

12.3

Scoring of graphical models

When we learn a graphical model from a set of cases, we want the model
that best describes the data. We want to express this goodness using a
171

single number, so that we can easily compare different models. We call such
a number a score.
Several different scoring measures have been proposed. The HUGIN API
provides the following scores:
The log-likelihood of the data given the model. This is simply the sum
of the log-likelihoods of the individual cases.
Akaikes Information Criterion (AIC): This is computed as l , where
l is the log-likelihood and is the number of free parameters in the
model.
The JeffreysSchwarz criterion, also called the Bayesian Information
Criterion (BIC): This is computed as l 21 log n, where l and are
defined as above, and n is the number of cases.
The log-likelihood score doesnt take model complexity into account, whereas the AIC and BIC scores do.
The following functions assume that case data has been specified, that domain is compiled (because inference will be performed), and that the junction tree potentials are up-to-date with respect to the node tables and their
models (if any). If domain is a LIMID, then each case must specify a valid
evidence scenario (see Section 9.1.3). Moreover, the network must not contain functional trails between nodes that are not real-valued function nodes.
h double t h domain get log likelihood (h domain t domain)

Get the log-likelihood of the case data with respect to the graphical model of
domain.2 This is computed using the current conditional probability tables.
If this function is called immediately after the EM algorithm has been run
(for example, using h domain learn tables(176) ), the log-likelihood will be
computed with respect to the final tables computed by the EM algorithm.
But the function can also be used without the EM algorithm.
h double t h domain get AIC (h domain t domain)

Get the AIC score of the case data with respect to the graphical model of
domain.
h double t h domain get BIC (h domain t domain)

Get the BIC score of the case data with respect to the graphical model of
domain.
2

Prior to version 6.5 of the HUGIN API, this function could only be used after the EM
algorithm had run, and the function returned the log-likelihood computed with respect to
the conditional probability tables before the final iteration of the EM algorithm i.e., not the
final tables.

172

12.4

Learning network structure

The algorithm used by HUGIN for learning the network structure is known
as the PC algorithm [36]. Domain knowledge (i.e., knowledge of which
edges to include or exclude, directions of the edges, or both) is taken into
account. Such knowledge is specified as a set of edge constraints (see Section 12.5).
An outline of the algorithm is as follows:
Case data is specified using the functions described in Section 12.1.
Alternatively, the cases can be loaded from a data file using h domain
parse cases(169) .
Statistical tests for conditional independence of pairs of nodes (X, Y)
given sets of other nodes SXY (with the size of SXY varying from 0 to 3)
are performed.
An undirected graph (called the skeleton) is constructed: X and Y are
connected with an edge if and only if (1) the edge is required by the
edge constraints, or (2) the edge is permitted by the edge constraints
and no conditional independence relation for (X, Y) given a set SXY
was found in the previous step.
Edges for which directions are specified by the edge constraints are
directed according to the constraints (unless the constraints impose
directed cycles or invalid directions).
Colliders (also known as v-structures) (i.e., edges directed at a common node) and derived directions are identified. Edges are directed
such that no directed cycles are created.
The previous step results in a partially directed graph. The remaining
edges are arbitrarily directed (one at a time, each edge directed is
followed by a step identifying derived directions).
x

h status t h domain learn structure (h domain t domain)

This function creates directed links (found by the PC algorithm) between
the nodes of domain which must contain only chance nodes and no edges.
Case data must be specified in advance using h domain parse cases(169) or
the functions described in Section 12.1. The learned network respects the
given edge constraints (see Section 12.5) unless the constraints impose
directed cyles or invalid directions (a discrete node is not allowed to have a
continuous parent).
The PC algorithm only determines the structure of the network. It does not
calculate the conditional probability tables. This can be done using the h
domain learn tables(176) function (see Section 12.6).
173

If a log-file has been specified (using h domain set log file(109) ), then a log
of the actions taken by the PC algorithm is produced. Such a log is useful for debugging and validation purposes (e.g., to determine which edge
directions were determined from data and which were selected at random).
The dependency tests calculate a test statistic which is asymptotically 2 distributed assuming (conditional) independence. If the test statistic is
large, the independence hypothesis is rejected; otherwise, it is accepted.
The probability of rejecting a true independence hypothesis is set using the
following function.
h status t h domain set significance level
(h domain t domain, h double t probability)

Set the significance level (i.e., the probability of rejecting a true independence hypothesis) to probability (a value between 0 and 1) for domain. The
default value is 0.05.
In general, increasing the significance level results in more edges, whereas
reducing it results in fewer edges. With fewer edges, the number of arbitrarily directed edges generally decreases.
Reducing the significance level also reduces the run time of h domain learn
structure.
h double t h domain get significance level (h domain t domain)

Retrieve the current significance level for domain.

Another way to reduce the run time of the PC algorithm is to exploit parallelization. The number of threads to use can be specified with the h domain
set concurrency level(21) function.3

12.5

Domain knowledge

Background knowledge about the domain can be used to constrain the set of
networks that can be learned. Such knowledge can be used by the learning
algorithm to resolve ambiguities (e.g., deciding the direction of an edge).
Domain knowledge can be knowledge of the direction of an edge, the presence or absence of an edge, or both.
The enumeration type h edge constraint t is introduced to represent the
set of possible items of knowledge about a particular edge a b. The possibilities are:
3

This is useful if the computer has multiple processors, a processor with multiple cores,
R
R
or a processor that supports Intel
Hyper-Threading Technology (such as an Intel
CoreTM
R
R


or an Intel Xeon processor). The concurrency level should be less than or equal to the
number of simultaneous threads supported by the computer.

174

 h constraint none indicates that no constraints are imposed on the

learning process. Unless any of the below constraints has been specified, h constraint none is assumed.
h constraint edge required indicates that an edge must be present, but
the direction of the edge is unspecified.
h constraint edge forbidden indicates that no edge is permitted.
h constraint forward edge required indicates that an edge is required,
and that it must be directed from a to b.
h constraint backward edge required indicates that an edge is required,
and that it must be directed from b to a.
h constraint forward edge forbidden indicates that, if an edge is present, it must be directed from b to a.
h constraint backward edge forbidden indicates that, if an edge is present, it must be directed from a to b.
Moreover, the constant h constraint error is used to denote error returns
from the h node get edge constraint function below.
x

h status t h node set edge constraint

(h node t a, h node t b, h edge constraint t constraint)
Specify constraint as the learning constraint for the edge a b. Note that
this also specifies a symmetric constraint for the edge b a (e.g., specifying h constraint forward edge required for a b also entails specifying h
constraint backward edge required for b a).

h edge constraint t h node get edge constraint

(h node t a, h node t b)
Retrieve the learning constraint specified for the edge a b. If an error
occurs, h constraint error is returned.

12.6

Learning conditional probability tables

Before learning of conditional probability tables can take place, the data set
and the set of nodes for which conditional probability distributions should
be learned must be specified. This set of nodes is specified as the nodes
having experience tables. Experience tables are created by the h node get
experience table(160) function, and they are deleted by the h table delete(73)
function.
175

h status t h domain learn tables (h domain t domain)

Learn the conditional probability tables for all nodes of domain having experience tables. The input to the learning procedure is the case data set specified using the functions described in Section 12.1. If domain is a LIMID,
then each case must specify a valid evidence scenario (see Section 9.1.3).
In general, the learning algorithm needs to do inference, so domain must be
compiled before h domain learn tables is called. If the computer has sufficient main memory, inference can be speeded up by saving the junction tree
tables to memory (using h domain save to memory(142) ) prior to the call of
h domain learn tables. Also, the junction tree potentials must be up-to-date
with respect to the node tables (i.e., the CPTs, the policies, and the utility
tables) and their models (if any), and the equilibrium must be sum with no
evidence incorporated. Moreover, the network must not contain functional
trails between nodes that are not real-valued function nodes.
As a sanity check, the function checks that case data has been entered, and
that learning has been enabled for at least one node (i.e., that at least one
node has an experience table containing at least one nonnegative count).
If successful, h domain learn tables updates the conditional probability table
and the experience table for each node of domain that has an experience table. Moreover, the inference engine is reset using the new conditional probability tables (see h domain reset inference engine(142) ). A retract-evidence
operation is implicitly performed as the final step of h domain learn tables.
If an error occurs, the state of the conditional probability tables and the
inference engine is unspecified.
The method used is the EM algorithm. If the experience count corresponding to some discrete parent configuration is zero, then h domain learn tables
computes the best (maximum likelihood) conditional probability distribution for that configuration. If the experience count is positive, then this
count is combined with the pre-specified contents of the conditional probability table to form prior experience (this is known as penalized EM):
If the node is discrete, then the conditional probability distribution is
multiplied by the experience count, resulting in a set of prior counts.
These counts are added to the counts derived from the data set.
If the node is continuous, then the joint weak marginal over the parents of the node is computed (based on the pre-specified conditional
probability tables).4 From this potential (and the experience count),
the prior probability distribution over the parameters of the CG distribution (Section 2.6) is computed. The details are too complex to
cover in this manual.
4
Because this computation does not involve evidence, the result is affected by the conditional probability tables of the node and all of its ancestors (but no other nodes).

176

If the experience count is negative, then learning is disabled for the corresponding parent state configuration.
For the discrete nodes, the starting point of the EM algorithm consists of the
pre-specified conditional probability tables. If no tables have been specified,
uniform distributions are assumed. Sometimes, it is desirable to enforce zeros in the joint probability distribution. This is done by specifying zeros in
the conditional probability tables for the configurations that should be impossible (i.e., have zero probability). However, note that presence of cases
in the data set which are impossible according to the initial joint distribution
will cause the learning operation to fail.
For the continuous nodes, a suitable starting point is computed from the
case data.
Example 12.3 The following code loads the Asia domain [25] and makes sure
that all nodes except E have experience tables. All entries of these experience tables
are then set to 0 (because we want to compute maximum likelihood estimates of
the conditional probability tables). Note that newly created experience tables are
already filled with zeros.
h_domain_t d = h_kb_load_domain ("Asia.hkb", NULL);
h_node_t E = h_domain_get_node_by_name (d, "E");
h_node_t n = h_domain_get_first_node (d);
for (; n != NULL; n = h_node_get_next (n))
if (n != E)
{
h_boolean_t b = h_node_has_experience_table (n);
h_table_t t = h_node_get_experience_table (n);
if (b)
{
h_number_t *data = h_table_get_data (t);
size_t k = h_table_get_size (t);
for (; k > 0; k--, data++)
*data = 0.0;
}
}
if (h_node_has_experience_table (E))
h_table_delete (h_node_get_experience_table (E));
Now we read and enter into the domain a file of cases (data file). This is done
using the h domain parse cases(169) function (see Example 13.25 for an appropriate definition of error handler). After having ensured that the domain is compiled,
we call h domain learn tables in order to learn conditional probability tables for all
nodes except E. [We assume that the correct conditional probability table has already been specified for E, and that the other conditional probability tables contain
nonzero values.]

177

h_domain_parse_cases
(d, data_file, error_handler, data_file);
if (!h_domain_is_compiled (d))
h_domain_compile (d);
h_domain_learn_tables (d);
The h domain learn tables operation will also update the experience tables with
the counts derived from the file of cases. These experience counts can then form
the basis for the sequential learning feature. (But note that if some parent state
configurations are absent from the data set, then the corresponding experience
counts will be zero.)

The EM algorithm performs a number of iterations. The main purpose of

each iteration is to produce an updated set of conditional probability tables.
This updated set of tables is either used as input to the next iteration, or it
is (part of) the final output of the EM algorithm.
The set of tables produced by an iteration of the EM algorithm is at least as
good as the set given at the start of the iteration. The measure of goodness
is the logarithm of the probability of the case data with respect to the probability distribution determined by the set of tables at the start of the iteration.
This quantity is known as the log-likelihood, and the EM algorithm attempts
to maximize this quantity.
If a log-file has been specified (using h domain set log file(109) ), then the
log-likelihood computed in each iteration of the EM algorithm is reported
to the log-file, as well as the log-likelihood, AIC, and BIC scores of the final
model. (See Section 12.3.)
The EM algorithm terminates when the relative difference between the loglikelihood for two successive iterations becomes sufficiently small. This criterion is controlled by the following function.
x

h status t h domain set log likelihood tolerance

(h domain t domain, h double t tolerance)
Specify that the EM algorithm used by h domain learn tables should terminate when the relative difference between the log-likelihood for two successive iterations becomes less than tolerance (which must be a positive number).
If this function is not called, a tolerance value of 104 is assumed by the EM
algorithm.

h double t h domain get log likelihood tolerance

(h domain t domain)
Retrieve the current setting of the log-likelihood tolerance for domain. If an
error occurs, a negative number is returned.
178

It is also possible to specify directly the maximum number of iterations performed.

x

h status t h domain set max number of em iterations

(h domain t domain, size t count)
Specify that the EM algorithm used by h domain learn tables should terminate when count number of iterations have been performed (if count is
positive). If count is zero, no limit on the number of iterations is imposed
(this is also the initial setting).

h count t h domain get max number of em iterations

(h domain t domain)
Retrieve the current setting for domain of the maximum number of iterations
for the EM algorithm used by h domain learn tables. If an error occurs, a
negative number is returned.
The EM algorithm terminates when at least one of the conditions described
above becomes true.

Learning CPTs in classes

If we have a runtime domain constructed from a class in an object-oriented
model using h class create domain(58) or h class create dbn domain(65) , then
several nodes in the runtime domain will typically be copies of the same
class node (i.e., the last element of their source lists will be a common class
node in the following, we shall refer to that class node as the source node).
Such nodes should be assigned the same conditional probability table by the
EM algorithm.
The following function can (only) be used when domain is a runtime domain constructed from some class in an object-oriented model (i.e., domain
must be a domain created by h class create domain(58) or h class create dbn
domain(65) it must not be a domain loaded from a file or a domain constructed in some other way).
x

h status t h domain learn class tables (h domain t domain)

For each node of domain (which must be a runtime domain) such that both
the node and its source node have experience tables,5 the conditional probability and experience tables of both nodes are learned/updated, and the tables of the domain node will be identical to those of its source node.6 Nodes
5
Unless experience tables have been deleted (or created), a runtime node has an experience table if and only if the corresponding source node has an experience table. Any other
kind of modifications to the runtime domain or the object-oriented model may produce unexpected results.
6
This implies that two runtime nodes with a common source node will get identical tables
(assuming all involved nodes have experience tables).

179

(in domain or in the object-oriented model) without experience tables will

not be modified.
Learning takes place in the object-oriented model (i.e., the conditional probability and experience tables of nodes in the object-oriented model are modified), but the inference part (the expectation step, or E-step for short)
of the EM algorithm takes place in the runtime domain. The results of the
E-step are combined to produce new conditional probability tables for the
nodes in the object-oriented model. These tables are then copied back to
the runtime domain so that they will be used in the next E-step. As the final
step of the EM algorithm, the conditional probability and experience tables
are copied from the object-oriented model to the runtime domain.
The initial contents of the experience tables of nodes in the object-oriented
model form the basis for the computation of prior counts. (See explanation concerning prior counts above.) However, in the current implementation, prior experience must not be specified for a continuous node in an
object-oriented model if the node has more than one runtime instance with
an experience table (that is, the experience tables for such nodes must not
contain positive values).
The contents of the updated experience tables reflect the fact that many
runtime nodes contribute to the learning of the same source node (i.e., the
experience counts will be higher than the number of cases in the data set).
Otherwise, everything specified for the h domain learn tables(176) function
also apply to the h domain learn class tables function.
Note that h domain learn class tables updates tables of nodes in the runtime domain as well as tables of nodes in the classes comprising the objectoriented model. In fact, the set of updated tables in the classes is typically
the desired outcome of calling the function.
The general procedure for learning class tables is as follows:
(1) Make sure that experience tables have been created for the set of
nodes in the object-oriented model for which EM learning is desired.
(2) Create a runtime domain. (If the source node corresponding to a runtime node has an experience table, then an experience table will automatically be created for the runtime node.)
(3) Enter case data into the runtime domain.
(4) Compile the runtime domain.
(5) Call h domain learn class tables on the runtime domain.

180

Chapter 13

The NET Language

When a belief network or LIMID model has been constructed using the
HUGIN API functions described in Chapter 2 and Chapter 3, it can be filed
for later use.
A non-object-oriented model (i.e., a domain) can be stored in a file using
a portable (but undocumented) binary format known as the HUGIN KB
format (see Section 2.10).
As an alternative to a binary format, a textual format can be used. As opposed to a binary format, a textual format has the advantage that it can be
read, modified, and even created from scratch by means of a standard text
editor.
The HUGIN system uses a special-purpose language called the NET language for textual specifications of belief networks and LIMID models,
object-oriented as well as non-object-oriented. The HUGIN API provides
functions to parse (Section 13.9) and generate (Section 13.10) text files
containing such specifications.
When new features are added to the HUGIN system, the NET language is
similarly extended. However, NET files generated by older versions of the
software can always be read by newer versions of the software. Also, the
NET language is extended in such a way that unless newer features are
being used in a NET file, then the file can also be read by an older version
of the HUGIN API (provided it is version 2 or newer1 ).

The first revision of the NET language (used by versions 1. x of the HUGIN API) had a
fixed format (i.e., the semantics of the different elements were determined by their positions
within the specification). This format could not (easily) be extended to support new features,
so a completely different format had to be developed.

181

13.1

Overview of the NET language

A domain or a class specification in the NET language is conceptually comprised of the following parts:
Information pertaining to the domain or class as a whole.
Specification of basic nodes (category, kind, states, label, etc).
Specification of the relationships between the nodes (i.e., the network
structure, and the potentials and functions associated with the links).
[Classes only] Specification of class instances, including bindings of
interface nodes.
[DBN classes only] Specification of special nodes, known as temporal
clones, to express temporal dependencies.
The first part (i.e., the part providing global information about a domain or
a class) must appear first, but the other parts can be overlapping, except that
nodes must be defined before they can be used in specifications of structure
or quantitative data.
A specification of a domain in the NET language has the following form:
hdomain definitioni hdomain headeri hdomain elementi*
hdomain headeri

net { hattributei* }

hdomain elementi hbasic nodei | hpotentiali

hattributei

hattribute namei = hattribute valuei ;

A specification of a class in the NET language has the following form:

hclass definitioni

class hclass namei { hclass elementi* }

hclass elementi

hdomain elementi | hattributei

| hclass instancei | htemporal clonei

A NET file can contain several class definitions. The only restriction is that
classes must be defined before they are instantiated.
Names (hclass namei, hattribute namei, etc.) are specified in Section 13.8.
The following sections describe the syntax and semantics of the remaining
elements of the grammar: hbasic nodei (Section 13.2), hclass instancei (Section 13.3), htemporal clonei (Section 13.4), and hpotentiali (Section 13.5
and Section 13.6),
182

13.2

Basic nodes

In ordinary belief networks, a node represents a random variable (either

discrete or continuous) this is known as a chance node. In LIMIDs, a node
can also represent a (discrete) decision, controlled by the decision maker, or
a utility function, which is used to assign preferences to different configurations of variables. Both types of networks can be extended with function
nodes. Function nodes can be used to perform computations using the values (or distributions) of other nodes as input. Function nodes can also be
used to provide input to the table generation process. There are two kinds of
function nodes: real-valued function nodes, representing a single real value,
and discrete function nodes, representing a discrete marginal distribution.
Evidence can be specified for discrete function nodes, but not for real-valued
function nodes.
This section explains how to specify these different types of nodes in a NET
specification. [In object-oriented models, nodes are also used to represent
class instances class instances are described in Section 13.3. And in dynamic models, special nodes, called temporal clones, are used to express
temporal dependencies see Section 13.4.]
Example 13.1 The following node specification is taken from the Chest Clinic
example [25].
node T
{
states = ("yes" "no");
label = "Has tuberculosis?";
position = (25 275);
}
This specifies a binary random variable named T , with states labeled "yes" and
"no". The specification also provides the label and position, which are used by the
HUGIN GUI tool.

A specification of a basic node begins with the specification of the node type:
[hprefixi] node (for specifying a chance node). The optional hprefixi
must be either discrete or continuous (omitting the hprefixi causes a
discrete chance node to be specified).
decision (for specifying a decision node).
utility (for specifying a utility node).
[discrete] function (for specifying a function node). If the optional
prefix discrete is included, the specification defines a discrete function
node. Otherwise, a real-valued function node is defined.
183

The node type specification is followed by a name that must be unique within the model. See Section 13.8 for the rules of forming valid node names.
A LIMID model (i.e., a model containing decision or utility nodes) must not
contain continuous (chance) nodes.
Example 13.1 shows three of the attributes currently defined in the NET
language for nodes: states, label, position, subtype, and state values. All of
these attributes are optional: If an attribute is absent, a default value is used.
states specifies the states of the node: This must be a non-empty list
of strings, comprising the labels of the states. If the node is used as
a labeled node with the table generator facility, then the labels must
be unique; otherwise, the labels need not be unique (and can even be
empty strings). The length of the list defines the number of states of
the node, which is the only quantity needed by the HUGIN inference
engine.
The default value is a list of length one, containing an empty string
(i.e., the node has only one state).
The states attribute is only allowed for discrete nodes.
label is a string that is used by the HUGIN GUI tool when displaying
the nodes. The label is not used by the inference engine. The default
value is the empty string.
position is a list of integers (the list must have length two). It indicates
the position within the graphical display of the network by the HUGIN
GUI tool. The position is not used by the inference engine. The default
position is at (0, 0).
subtype specifies the subtype of a discrete node. The value must be
one of the following name tokens: label, boolean, number, or interval.
See Section 6.1 for more information.
The default value is label.
state values is a list of numbers, defining the state values of the node.
These values are used by the table generator facility. This attribute
must only appear for nodes of subtypes number or interval (and must
appear after the subtype and states attributes). If the subtype is number, the list must have the same length as the states list; if the subtype
is interval, the list must have one more element than the states list.
The list of numbers must form an increasing sequence.
If the subtype is interval, the first element can be infinity, and the
last element can be infinity.
In addition to the standard attributes, an application can introduce its own
attributes.
184

Example 13.2 Here, the node T has been given the application-specific attribute
MY APPL my attr.
node T
{
states = ("yes" "no");
label = "Has tuberculosis?";
position = (25 275);
MY_APPL_my_attr = "1000";
}

The values of such application-specific attributes must be text strings (see

Section 13.8 for a precise definition of text strings).
The names of application-specific attributes should have a common prefix
in order to avoid name clashes with attributes defined by HUGIN or other
applications (in Example 13.2 the MY APPL prefix is used).
When a NET specification has been parsed, the values of application-specific
attributes can be accessed using the h node get attribute(46) and h node set
attribute(46) functions.
Example 13.3 In the HUGIN GUI tool, some extra attributes are used to save descriptions of both nodes and their states. These are the attributes prefixed with HR .
node T
{
states = ("yes" "no");
label = "Has tuberculosis?";
position = (25 275);
HR_State_0 = "The patient has tuberculosis.";
HR_State_1 = "The patient does not have\
tuberculosis.";
HR_Desc = "Represents whether the patient has\
tuberculosis.";
}
The HUGIN GUI tool uses the UTF-8 encoding for storing arbitrary text in attributes.
If you need to have your networks loaded within that tool, you should use the UTF-8
encoding for non-ASCII text.

13.3

Class instances

In object-oriented models, a node can also represent a class instance. Such a

node is introduced using the instance keyword:
instance hinstance namei : hclass namei
( [hinput bindingsi] [ ; houtput bindingsi] ) { hnode attributesi }
185

This defines an instance (with name hinstance namei) of the class with name
hclass namei. Currently, the hnode attributesi for a class instance can only
contain label, position, and user-defined attributes.
The hinput bindingsi specify how formal input nodes of the class instance
are associated with actual input nodes. The syntax is as follows:
hinput bindingsi hinput bindingi , hinput bindingsi
hinput bindingi hformal input namei = hactual input namei
The hformal input namei must refer to a node listed in the inputs attribute
(see Section 13.7) of the class with name hclass namei. The node referred
to by the hactual input namei must be defined somewhere in the class containing the class instance.
The hinput bindingsi need not specify bindings for all the formal input nodes
of the class (but at most one binding can be specified for each input node).
The houtput bindingsi are used to give names to output clones. The syntax
is similar to that of the input bindings:
houtput bindingsi houtput bindingi , houtput bindingsi
houtput bindingi hactual output namei = hformal output namei
The hactual output namei is the name assigned to the output clone that corresponds to the output node with name hformal output namei for this particular class instance. An hactual output namei may appear in the outputs attribute (see Section 13.7) of a class definition and as a parent in hpotentiali
specifications.
Example 13.4 The following fragment of a NET specification defines an instance I1
of class C.
instance I1 : C (X=X1, Y=Y1; Z1=Z) {...}
Class C must have (at least) two input nodes: X and Y. For instance I1 , X corresponds to node X1 , and Y corresponds to node Y1 . Class C must also have (at least)
one output node: Z. The output clone corresponding to Z for instance I1 is given
the name Z1 .

A NET file can contain several class definitions, but the classes must be
ordered such that instantiations of a class follow its definition. Often, a
NET file will be self-contained (i.e., no class instances refer to classes not
defined in the file), but it is also possible to store the classes in individual
files. When a NET file is parsed, classes will be looked up whenever they
are instantiated. If the class is already loaded, the loaded class will be used.
If no such class is known, it must be created (for example, by calling the
parser recursively). See Section 13.9 for further details.
186

13.4

Temporal clones

In dynamic models, special nodes, known as temporal clones, are used to

express temporal dependencies: If a node X at time t depends on the value
(or state) of X at time t 1, then this dependency can be expressed by a link
from a temporal clone of X to X.
Temporal clones are introduced by the temporal keyword.
temporal hclone namei : hmaster namei { hnode attributesi }
This defines a temporal clone (with name hclone namei) of the node with
name hmaster namei. The definition of the master node must appear before the definition of the temporal clone. The master node must not be an
instance node, an output clone, or a temporal clone.
At most one temporal clone can be specified for a given master node.
Temporal clones can only be defined in classes (not domains). These classes
must not have input or output nodes, and they must not be instantiated in
other classes.
The hnode attributesi for a temporal clone can only contain label, position,
and user-defined attributes.
The category and kind of the temporal clone are inherited from the master
node. If the master node is discrete, then the temporal clone has the same
subtype, the same number of states, the same state labels, and the same
state values as the master node.

13.5

The structure of the model

The structure (i.e., the links of the underlying graph) is specified as part
of the hpotentiali specifications. We have two kinds of links: directed and
undirected links. We denote a directed link from A to B as A B, and we
denote an undirected link between A and B as A B. If there is a directed
link from A to B, we say that A is a parent of B and that B is a child of A.
A network model containing undirected links is called a chain graph model.
A hpotentiali specification is introduced by the potential keyword. In the
following, we explain how to specify links through a series of examples.
Example 13.5 This is a typical specification of directed links:
potential ( A | B C ) { }
This specifies that node A has two parents: B and C. That is, there is a directed
link from B to A, and there is also a directed link from C to A.
Example 13.6 This is a typical specification of undirected links:

187

potential ( A B ) { }
This specifies that there is an undirected link between A and B. (If there are no
parents, the vertical bar may be omitted.)
Example 13.7 Directed and undirected links can also be specified together:
potential ( A B | C D ) { }
This specifies that there is an undirected link between A and B, and it also specifies
that A and B both have C and D as parents.

A maximal set of nodes, connected by undirected links, is called a chain

graph component. If a chain graph component has more than one member,
then all members must be discrete chance nodes (i.e., undirected links are
only allowed between discrete chance nodes).
The graph must not contain a directed cycle: A directed cycle is a sequence
of nodes X1 , . . . , Xn (n > 1 and X1 = Xn ), such that either Xi Xi+1 or
Xi Xi+1 (1 i < n), and at least one of the links is directed.
Example 13.8 This specification is invalid, because of the directed cycle A B
C A.
potential ( B | A ) { }
potential ( C | B ) { }
potential ( A | C ) { }
If we reverse the link from C to A, we obtain a valid specification.
potential ( B | A ) { }
potential ( C | A B ) { }

Example 13.9 This specification is also invalid, since there is a directed cycle A
B C A.
potential ( B | A ) { }
potential ( C | B ) { }
potential ( A C ) { }
However, the following specification is valid.
potential ( A | B ) { }
potential ( C | B ) { }
potential ( A C ) { }

If there is only one child node in the hpotentiali specification, then this specifies the node table associated with the child node (Section 13.6 explains
188

how to specify the numeric parts of the table). Only one such specification
can be provided for each node.
Chance, decision, and utility nodes can only have chance, decision, and
function nodes as parents. Discrete (chance and decision) nodes cannot
have continuous nodes as parents (but discrete function nodes can).
In dynamic models, temporal clones can only have temporal clones as parents (that is, links must follow the natural flow of time).
Links entering function nodes and links leaving real-valued function nodes
are called functional links. Additional constraints are imposed on network
models containing functional links. See Section 2.4.
Links entering decision nodes are called information links. The parents of a
decision node are exactly those nodes that are assumed to be known when
the decision is made.
Example 13.10 Assume that we want to specify a LIMID with two decisions, D1
and D2 , and with three discrete chance variables, A, B, and C. First, A is observed;
then, decision D1 is made; then, B is observed; finally, decision D2 is made. This
sequence of events can be specified as follows:
potential ( D1 | A ) { }
potential ( D2 | D1 B ) { }
The last line specifies that the decision maker forgets the observation of A before
he makes decision D2 . If this is not desired, then A should be included in the
hpotentiali specification for D2 :
potential ( D2 | D1 A B ) { }
However, this makes the policy of D2 more complex. Reducing the comlexity of
decision making by ignoring less important observations can often be an acceptable
(or even necessary) trade-off.

Finally, a node must not be referenced in a hpotentiali specification before it

has been defined by a node-, decision-, utility-, function-, or a temporalspecification.

13.6

Potentials

We also need to specify the quantitative part of the model. This part consists of conditional probability potentials for chance and discrete function
nodes, policies for decision nodes, and utility functions for utility nodes. We
distinguish between discrete probability, continuous probability, and utility
potentials. Discrete probability potentials are used for all discrete nodes
(including decision nodes).
189

There are two ways to specify the discrete probability and the utility potentials: (1) by listing the numbers comprising the potentials (Section 13.6.1),
and (2) by using the table generation facility (Section 13.6.2).
Real-valued function nodes do not have potentials, but it is convenient to use
hpotentiali specifications to define the functions represented by the nodes.
In this case, we use expressions to specify the functions (see Section 13.6.2).

13.6.1

Direct specification of the numbers

Direct specification of the quantitative part of the relationship between a

group of nodes and their parents is done using the data attribute of the
hpotentiali specification.
Example 13.11 This specification is taken from the Chest Clinic example [25]. It
specifies the conditional probability table of the discrete variable T (tuberculosis).
potential ( T | A )
{
data = (( 0.05 0.95 )
( 0.01 0.99 ));
}

%
%

A=yes
A=no

This specifies that the probability of tuberculosis given a trip to Asia is 5%, whereas
it is only 1% if the patient has not been to Asia.
The data attribute may also be specified as an unstructured list of numbers:
potential ( T | A )
{
data = ( 0.05 0.95
0.01 0.99 );
}

%
%

A=yes
A=no

As the example shows, the numerical data of a discrete probability potential

is specified through the data attribute of a hpotentiali specification. This
data has the form of a list of real numbers.
For chance and decision nodes, the structure of the list must either correspond to that of a multi-dimensional table with dimension list comprised
of the (discrete) parent nodes followed by the child nodes, or it must be a
flat list with no structure at all. In both cases, the layout of the data list is
row-major (see Section 5.1).
For discrete function nodes, only the node itself is included in the probability table, so the attribute value is simply a vector of numbers. However,
the probability table of a discrete function node is usually generated from a
model see Section 13.6.2.
190

Example 13.12
potential ( D E F | A B C ) { }
The data attribute of this hpotentiali specification corresponds to a multi-dimensional table with dimension list hA, B, C, D, E, Fi.

The data attribute of a utility potential corresponds to a multi-dimensional

table with dimension list comprised of the (discrete) nodes to the right of
the vertical bar.
Example 13.13 This specification is taken from the Oil Wildcatter example and
shows a utility potential: DrillProfit is a utility node, Oil is a discrete chance
node with three states, and Drill is a decision node with two states.
potential (DrillProfit | Drill Oil)
{
data = (( -70
% Drill=yes
50
% Drill=yes
200 )
% Drill=yes
(
0
% Drill=no
0
% Drill=no
0 ));
% Drill=no
}

Oil=dry
Oil=wet
Oil=soaking
Oil=dry
Oil=wet
Oil=soaking

The data attribute of this hpotentiali specification corresponds to a multi-dimensional table with dimension list hDrill, Oili.

The (multi-dimensional) table corresponding to the data attribute of a continuous probability potential has dimension list comprised of the discrete
parents of the hpotentiali specification (in the given order). These nodes
must be listed first on the right hand side of the vertical bar, followed by the
continuous parents. However, the items in the table are no longer numbers
but instead continuous distribution functions; only normal (i.e., Gaussian)
distributions can be used. A normal distribution is specified by its mean and
variance. In the following example, a continuous probability potential is
specified.
Example 13.14 Suppose A is a continuous node with parents B and C, which are
both discrete. Also, both B and C have two states: B has states b1 and b2 while C
has states c1 and c2 .
potential (A | B C)
{
data = (( normal
normal
( normal
normal
}

(
0,
( -1,
(
1,
( 2.5,

191

1
1
1
1.5

)
) )
)
) ));

%
%
%
%

B=b1
B=b1
B=b2
B=b2

C=c1
C=c2
C=c1
C=c2

The data attribute of this hpotentiali specification corresponds to a table with dimension list hB, Ci. Each entry contains a probability distribution for the continuous node A.

All entries in the above example contain a specification of a normal distribution. A normal distribution is specified using the keyword normal followed
by a list of two parameters. The first parameter is the mean and the second
parameter is the variance of the normal distribution.
Example 13.15 Let A be a continuous node with one discrete parent B (with states
b1 and b2 ) and one continuous parent C.
potential (A | B C)
{
data = ( normal ( 1 + C, 1 )
normal ( 1 + 1.5 * C, 2.5 ) );
}

%
%

B=b1
B=b2

The data attribute of this hpotentiali specification corresponds to a table with dimension list hBi (B is the only discrete parent, and it must therefore be listed first
on the right hand side of the vertical bar). Each entry again contains a continuous
distribution function for A. The influence of C on A now comes from the use of C
in the expressions specifying the mean parameters of the normal distributions.

Only the mean parameter of a normal distribution can be specified as an

expression. The variance parameter must be a numeric constant. The expression for the mean parameter must be a linear function of the continuous
parents: each term of the expression must be (1) a numeric constant, (2) the
name of a continuous parent, or (3) a numeric constant followed by * followed by the name of a continuous parent.
If the data attribute of a hpotentiali specification is missing, a list of 1s is
assumed for discrete probability potentials, whereas a list of 0s is assumed
for utility potentials. For a continuous probability potential, a list of normal
distributions with mean set to 0 and variance set to 1 is assumed.
The values of the data attribute of discrete probability potentials must contain only nonnegative numbers. In the specification of a normal distribution
for a continuous probability potential, only nonnegative numbers are allowed for the variance parameter. There is no such restriction on the values
of utility potentials or the mean parameter of a normal distribution.

13.6.2

Using the table generation facility

For potentials that do not involve CG variables, a different method for specifying the quantitative part of the relationship for a single node and its parents is provided. This method can be used for all discrete nodes as well as
utility nodes.
192

Models are also used to define the functions associated with real-valued
function nodes. This is explained below.
Example 13.16 Let A denote the number of 1s in a throw with B (possibly biased)
dice, where the probability of getting a 1 in a throw with a single die is C. The
specification of the conditional probability potential for A given B and C can be
given using the table generation facility described in Chapter 6 as follows:
potential (A | B C)
{
model_nodes = ();
samples_per_interval = 50;
model_data = ( Binomial (B, C) );
}
First, we list the model nodes attribute: This defines the set of configurations for the
model data attribute. In this case, the list is empty, meaning that there is just one
configuration. The expression for that configuration is the binomial distribution
expression shown in the model data attribute.
C will typically be an interval node (i.e., its states represent intervals). However,
when computing the binomial distribution, a specific value for C is needed. This
is handled by choosing 50 distinct values within the given interval and computing
the distributions corresponding to those values. The average of these distributions
is then taken as the conditional distribution for A given the value of B and the
interval (i.e., state) of C. The number 50 is specified by the samples per interval
attribute. See Section 6.9 for further details.
Example 13.17 In the Chest Clinic example [25], the node E is specified as a
logical OR of its parents, T and L. Assuming that all three nodes are of labeled
subtype with states yes and no (in that order), the potential for E can be specified
as follows:
potential (E | T L)
{
model_nodes = (T L);
model_data = ( "yes", "yes", "yes", "no" );
}
An equivalent specification can be given in terms of the OR operator:
potential (E | T L)
{
model_nodes = ();
model_data
= ( if (or (T="yes", L="yes"), "yes", "no") );
}
If all three nodes are given a boolean subtype, the specification can be simplified to
the following:

193

potential (E | T L)
{
model_nodes = ();
model_data = ( or (T, L) );
}

In general, the model nodes attribute is a list containing a subset of the discrete parents listed to the right of the vertical bar in the hpotentiali specification. The order of the nodes in the model nodes list defines the interpretation
of the model data attribute: The model data attribute is a comma-separated
list of expressions, one for each configuration of the nodes in the model
nodes list. As usual, the ordering of these configurations is row-major.
A non-empty model nodes list is a convenient way to specify a model with
distinct expressions for distinct parent state configurations. An alternative
is nested if-expressions. See Example 13.17.
The model nodes attribute must appear before the samples per interval and
model data attributes.
The complete definition of the syntax of expressions is given in Section 6.3.
If both a specification using the model attributes and a specification using
the data attribute are provided, then the specification in the data attribute
is assumed to be correct (regardless of whether it was generated from the
model). The functions that generate NET files (Section 13.10) will output
both, if HUGIN believes that the table is up-to-date with respect to the
model (see the description of h node generate table(95) for precise details).
Since generating a table from its model can be a very expensive operation,
having a (redundant) specification in the data attribute can be considered a
cache for h node generate table.
Real-valued function nodes
Real-valued function nodes do not have potentials, but it is convenient to use
expressions to specify the functional relationship between a (real-valued)
function node and its parents.
This is achieved by using a hpotentiali specification containing only the
model nodes and model data attributes (the data attribute is not allowed).
In this case, all expressions must be of numeric type.
Example 13.18 Real-valued function nodes can be used to provide important constants in expressions. This is done by introducing a (real-valued) function node
(named appropriately) and defining the constant using an expression.
If we would like to fix the value of C in Example 13.16 to 0.2, we could define C as
a (real-valued) function node and define its value using a hpotentiali specification:

194

potential (C)
{
model_nodes = ();
model_data = ( 0.2 );
}
Defining the probability parameter of the binomial distribution this way makes it
easy to change it later.

If real-valued function (or utility) nodes are used as parents in a hpotentiali

specification, then they must appear last in the list of parents.

13.6.3

Parameter learning

Information for use by the adaptation facility (see Chapter 11) is specified
through the experience and fading attributes of a hpotentiali specification.
These attributes have the same syntax as the data attribute.
The experience data is also used to control the EM algoritm (Section 12.6).
The experience attribute is only allowed in hpotentiali specifications for single (that is, there must be exactly one child node in the specification)
chance nodes. The fading attribute is only allowed in hpotentiali specifications for single discrete chance nodes.
For the adaptation algoritm, valid experience counts must be positive numbers, while the EM algoritm only requires nonnegative numbers. Specifying
an invalid value for some parent state configuration turns off parameter
learning for that configuration. If the hpotentiali specification doesnt contain an experience attribute, parameter learning is turned off completely for
the child node.
A fading factor is valid if 0 < 1. Specifying an invalid fading factor for
some parent state configuration turns off adaptation for that configuration.
If the hpotentiali specification doesnt contain a fading attribute, then all
fading factors are considered to be equal to 1 (which implies no fading).
See Chapter 11 and Section 12.6 for further details.
Example 13.19 The following shows a specification of experience and fading information for the node D (Dyspnoea) from the Chest Clinic example in [25].
This node has two parents, E and B. We specify an experience count and a fading
factor for each configuration of states of hE, Bi.
potential (D | E B)
{
data = ((( 0.9 0.1
( 0.7 0.3
(( 0.8 0.2
( 0.1 0.9
experience = (( 10

)
))
)
)));

%
%
%
%
%

195

E=yes
E=yes
E=no
E=no
E=yes

B=yes
B=no
B=yes
B=no
B=yes

12 )
0
14 ));

fading = (( 1.0
0.9 )
( 1.0
1.0 ));

%
%
%
%
%
%
%

E=yes
E=no
E=no
E=yes
E=yes
E=no
E=no

B=no
B=yes
B=no
B=yes
B=no
B=yes
B=no

}
Note that the experience count for the E = no/B = yes state configuration is 0. This
value is an invalid experience count for the adaptation algoritm (but not for the EM
algoritm), so adaptation is turned off for that particular state configuration. Also,
note that only the experience count for the E = yes/B = no state configuration will
be faded during adaptation (since the other parent state configurations have fading
factors equal to 1).

13.7

Global information

Information pertaining to the belief network or LIMID model as a whole is

specified as attributes within the hdomain headeri (for domains) or within
the hclass definitioni (for classes).
Example 13.20 The HUGIN GUI tool uses several parameters when displaying networks.
net
{
node_size = (100 40);
}
This specifies that nodes should be displayed with width 100 and height 40.

Currently, only the node size attribute is recognized as a special global attribute. However, as with nodes, extra attributes can be specified. These
extra attributes must take strings as values. The attributes are accessed
using the HUGIN API functions h domain get attribute(46) , h domain set attribute(46) , h class get attribute(62) , and h class set attribute(61) .
The HUGIN GUI tool uses the UTF-8 encoding for storing arbitrary text in
attributes. If you need to have your networks loaded within that tool, you
should use the UTF-8 encoding for non-ASCII text.
Example 13.21
net
{
node_size = (100 40);
MY_APPL_my_attr = "1000";
}
This specification has an application specific attribute named MY APPL my attr.

196

Example 13.22 Recent versions of the HUGIN GUI tool use several application
specific attributes. Some of them are shown here:
net
{
node_size = (80 40);
HR_Grid_X = "10";
HR_Grid_Y = "10";
HR_Grid_GridSnap = "1";
HR_Grid_GridShow = "0";
HR_Font_Name = "Arial";
HR_Font_Size = "-12";
HR_Font_Weight = "400";
HR_Font_Italic = "0";
HR_Propagate_Auto = "0";
}
HUGIN GUI uses the prefix HR on all of its application specific attributes (a predecessor of the HUGIN GUI tool was named HUGIN Runtime).

Global attributes are used in a hclass definitioni to specify the interface of

the class. The inputs and outputs attributes are used to specify the input
nodes and the output nodes of the class, respectively. The values of these
attributes are node lists (with the same syntax as that of the model nodes
attribute). The nodes mentioned in those attributes must be defined within
the class.
Example 13.23 The following class specification defines a class C with two inputs,
X and Y, and one output, Z.
class C
{
inputs = (X Y);
outputs = (Z);
node X
...
node Y
...
node Z
...
...
}

13.8

Lexical matters

In general, a name has the same structure as an identifier in the C programming language. That is, a name is a non-empty sequence of letters and
197

digits, beginning with a letter. In this context, the underscore character ( )

is considered a letter. The case of letters is significant. The sequence of letters and digits forming a name extends as far as possible; it is terminated by
the first non-letter/digit character (for example, braces or whitespace).
As an exception, a node name in a domain specification (but not a class
specification) can be a dotted name: A list of names (as specified above)
with a single dot (a period) separating each pair of consecutive names in the
list. (Note: Whitespace is not allowed in a dotted name.) This exception
makes it possible to assign unique meaningful names to nodes in a runtime
domain (see Section 3.10).
A string is a sequence of characters not containing a quote character (") or
a newline character; its start and ending are indicated by quote characters.
A number is comprised of an optional sign, followed by a sequence of digits,
possibly containing a decimal point character, and an optional exponent
field containing an E or e followed by an (optionally signed) integer.
Comments can be placed in a NET specification anywhere (except within a
name, a number, or other multi-character lexical elements). It is considered
equivalent to whitespace. A comment is introduced by a percent character
(%) and extends to the end of the line.

13.9

Parsing NET files

The HUGIN API provides two functions for parsing models specified in the
NET language: one for non-object-oriented models (domains), and one
for object-oriented models (classes).
This function parses non-object-oriented specifications (i.e., NET files starting with the net keyword) and creates a corresponding h domain t object.
x

h domain t h net parse domain

(h string t file name,
void (error handler) (h location t, h string t, void ),
void data)
Parse the NET specification in the file with name file name. If an error is
detected (or a warning is issued), the error handler function is called with
a line number (indicating the location of the error within the file), a string
that describes the error, and data. The storage used to hold the string is
reclaimed by h net parse domain when error handler returns (so if the error/warning message will be needed later, a copy must be made).
The user-specified data allows the error handler to access non-local data
(and hence preserve state between calls) without using global variables.
The h location t type is an unsigned integer type (such as unsigned long).
198

If no error reports are desired (in this case, only the error indicator returned
by h error code(18) will be available), then the error handler argument may
be NULL. (In this case, warnings will be completely ignored.)
If the NET specification is successfully parsed, an opaque reference to the
created domain structure is returned; otherwise, NULL is returned. The domain is not compiled; use a compilation function to get a compiled version.
Example 13.24 The error handler function could be written as follows.
void my_error_handler
(h_location_t line_no, h_string_t message, void *data)
{
fprintf (stderr, "Error at line %d: %s\n",
line_no, message);
}
This error handler simply writes all messages to stderr. See Example 13.25 for a
different error handler.

The following function must be used when parsing NET files containing class
specifications (i.e., NET files starting with the class keyword).
x

h status t h net parse classes

(h string t file name, h class collection t cc,
void (get class) (h string t, h class collection t, void ),
void (error handler) (h location t, h string t, void ),
void data)
This function parses the contents of the file with name file name. This file
must contain a sequence of class definitions. The parsed classes are stored
in class collection cc.
In order to create the instance nodes (which represent instances of other
classes), it may be necessary to load these other classes: If an instance of a
class not present in cc is specified in the NET file, get class is called with the
name of the class, the class collection cc, and the user-specified data. The
get class function is supposed to load the named class into class collection cc
(if it doesnt, then parsing is terminated). If the named class contains instances of other classes not present in cc, these classes must be loaded (or
constructed) as well. The get class function should not perform any other
kind of actions. For example, it should not delete or rename any of the
existing classes in cc such actions may cause the HUGIN API to crash.
If the specified NET file is self-contained (i.e., no instance declaration refers
to a class not specified in the file), then the get class argument can be NULL.
Note that instance nodes are created when the corresponding instance definition is seen in the NET file. At that point, the instantiated class must
have been loaded (or get class will be called). For this reason, if the NET file
199

contains several class definitions, classes must be defined before they are
instantiated.
If an error is detected, the error handler function is called with a line number, indicating the location within the source file currently being parsed,
and a string that describes the error. The storage used to hold this string is
reclaimed by h net parse classes when error handler returns (so if the error
message will be needed later, a copy must be made).
If parsing fails, then h net parse classes will try to preserve the initial contents of cc by deleting the new (and possibly incomplete) classes before it
returns. If get class has modified any of the classes initially in cc, then this
may not be possible. Also, if the changes are sufficiently vicious, then removing the new classes might not even be possible. However, if get class
only does things it is supposed to do, there will be no problems.
As described above, the get class function must insert a class with the specified name into the given class collection. This can be done by whatever
means are convenient, such as calling the parser recursively, or through explicit construction of the class.
Example 13.25 Suppose we have classes stored in separate files in a common directory, and that the name of each file is the name of the class stored in the file with
.net appended. Then the get class function could be written as follows:
void get_class
(h_string_t name, h_class_collection_t cc, void *data)
{
h_string_t file_name = malloc (strlen (name) + 5);
if (file_name == NULL)
return;
(void) strcat (strcpy (file_name, name), ".net");
(void) h_net_parse_classes
(file_name, cc, get_class, error_handler,
file_name);
free (file_name);
}
void error_handler
(h_location_t line_no, h_string_t err_msg, void *data)
{
fprintf (stderr, "Error in file %s at line %lu: %s\n",
(h_string_t) data, (unsigned long) line_no,
err_msg);
}

200

Note that we pass the file name as the data argument to h net parse classes. This
means that the error handler receives the name of the file as its third argument.
If more data is needed by either get class or the error handler, then the data argument can be specified as a pointer to a structure containing the needed data
items.

13.10

Saving class collections, classes, and domains

as NET files

The following functions can be used to create NET files.

x

h status t h cc save as net

(h class collection t cc, h string t file name)

h status t h class save as net (h class t class, h string t file name)

h status t h domain save as net

(h domain t domain, h string t file name)
Save the class collection, class, or domain as a text file with name file name.
The format of the file is as required by the NET language.
Saving a class collection as a NET file is convenient when you must send the
object-oriented model via email, since the resulting NET specification must
necessarily be self-contained.
Note that if a NET file is parsed and then saved again, any comments in
the original file will be lost. Also note that if (some of) the nodes have not
been assigned names, then names will automatically be assigned (through
calls to the h node get name(42) function). Likewise, if a class has not been
named, the save-as-NET operation will assign a name (by calling h class
get name(53) ).

h string t h class get file name (h class t class)

Return the file name used for the most recent (successful) save-operation
applied to class. If no such operation has been performed, or class is NULL,
NULL is returned.

h string t h domain get file name (h domain t domain)

Return the file name used for the most recent (successful) save-operation
applied to domain. If no such operation has been performed, or domain is
NULL , NULL is returned.
Note that domains may be saved both as a NET and as a HUGIN KB file.

201

202

Chapter 14

Data Sets and CSV Files

Data is not always provided in a format that can be readily loaded by the
HUGIN API facilities described in Section 12.2. Often data is provided as
a so-called CSV file (short for comma-separated-values file). There are
many similarities between CSV files and Hugin data files, but there are also
some differences. The data in a CSV file is basically untyped, whereas the
HUGIN API functions for loading a Hugin data file impose restrictions on
the data items that are acceptable. The acceptable data items are defined by
the state spaces of the nodes that correspond to the columns (i.e., the positions of fields within the lines) of the data file. These restrictions allow the
functions for loading Hugin data files to be very memory efficient, because
the data is entered directly as case data during the parsing process. A CSV
file, on the other hand, is loaded as pure text. Because of that, the memory
consumption is higher than the size of the file (as there is some overhead to
the representation of the data).
The HUGIN API provides functions to aid in the loading of data provided
as CSV files. A data type to hold this data internally in the memory of a
computer is also provided. This data type is called a data set.

14.1

Data sets

A data set provides a row/column (or a matrix) representation of text values,

with rows as cases and columns as variables. The columns of a data set are
named these names will typically be used for naming nodes associated
with the coulumns.
A data set can be loaded from a CSV file, and it can be written as (i.e., saved
to) a CSV file. The first line of the CSV file is assumed to contain the names
of the columns. Functions are provided for building data sets and for manipulating the text data items of a data set. A function is also provided for
adding the rows of a data set as case data for a domain.
203

Data sets are represented as C objects of type h data set t.

x

h data set t h new data set (void)

Create a new (empty) data set.

h status t h ds delete (h data set t data set)

Deallocate all storage associated with data set.

h index t h ds new column (h data set t data set, h string t name)

Create a new (empty) column in data set named name. The name can be
any text string, but if it should later be used as the name of a node, then the
usual requirements for node names must be satisfied.
The index associated with the new column is returned. If an error occurs, a
negative number is returned.

h index t h ds new row (h data set t data set)

Create a new (empty) row in data set. The index associated with the new
row is returned. If an error occurs, a negative number is returned.

h status t h ds delete column (h data set t data set, size t column)

Delete the specified column from data set. Columns to the right of the
deleted column are shifted one position to the left.1

h status t h ds delete row (h data set t data set, size t row)

Delete the specified row from data set. Rows below the deleted row are
shifted one position up.1

h status t h ds set column name

(h data set t data set, size t column, h string t name)
Change the name of the specified column of data set to name.

h string t h ds get column name

(h data set t data set, size t column)
Return the name of the specified column of data set.

h count t h ds get number of columns (h data set t data set)

Return the number of columns in data set.

h count t h ds get number of rows (h data set t data set)

Return the number of rows in data set.
1
Data sets are usually visualized as tables with columns numbered from left to right and
rows numbered from top to bottom.

204

h status t h ds move column

(h data set t data set, size t column, size t new column)
Move the specified column to a new position. The columns between the old
and the new column positions are shifted one position to the left or to the
right depending on the direction of the move.

h status t h ds move row

(h data set t data set, size t row, size t new row)
Move the specified row to a new position. The rows between the old and
the new row positions are shifted one position up or down depending on
the direction of the move.

h status t h ds set data item

(h data set t data set, size t row, size t column, h string t data)
Set (or delete) the data item at the specified location of data set. If data is
NULL , then the value at the specified location is deleted (indicating a missing data item).

h string t h ds get data item

(h data set t data set, size t row, size t column)
Return the data item stored at the specified location of data set. If there is
no data item (i.e., the data item is missing) at the specified location, NULL
is returned.
Note that the text string returned by this function is the actual data string
stored in data set. This string must not be freed.

h status t h domain add cases

(h domain t domain, h data set t data set, size t first row,
size t number of rows)
The range of rows of data set, starting at index first row and ending at index
first row + number of rows 1, are added as cases to domain. The names
of nodes in domain and columns in data set are used to match nodes with
columns. Columns in data set with no matching nodes in domain are ignored. When a node has been matched to a column, the data in the column
is entered as case data for the node. As a special case, a column named #
is treated as containing case counts see h domain set case count(167) .
If the data cant be converted to appropriate state information, the function
fails (and no data is entered).

205

14.2

CSV files

CSV files (short for comma-separated-values files) are used as external

representations of data sets. With that purpose in mind, the HUGIN API
accepts CSV files that conform to the following format:
Each row of the data set is represented by a single line in the file. The
individual data items in a row are represented by fields within the
line separated by commas (some other character, such as a colon or a
semicolon, can also be used for separation).2
The first (non-empty) line defines the names of all columns of the data
set. The number of fields in that line defines the number of columns
of the data set. The number of fields in each remaining line must not
exceed the number of columns in the data set. Fewer fields in a line
indicate missing values of the righmost columns.
Double-quotes must be used to delimit fields containing the chosen
delimiter or a double-quote character. Each double-quote character
within the data item must be represented by two double-quotes. The
initial and final double-quotes are not part of the data item itself. The
double-quote character cannot be used as a delimiter.
Initial and trailing whitespace of a field is ignored. Data items that
contain initial or trailing whitespace must be quoted.
Empty lines (except for optional whitespace) are ignored (that is, they
do not define empty rows in the data set).
h data set t h csv parse data set
(h string t file name, int delimiter,
void (error handler) (h location t, h string t, void ),
void data)

Parse a CSV file that uses delimiter to separate the fields of a record (line).
The delimiter should be a normal character (that is, it should be visible
and not a so-called control character). However, a blank or a tab (but not
a double-quote) character can be used as delimiter.
The error handler and data arguments are used for error handling. This is
similar to the error handling done by the other parse functions. See Section 13.9 for further information.
h status t h ds save
(h data set t data set, h string t file name, int delimiter)

Save data set in the format of a comma-separated-values (CSV) file. However, another delimiter than a comma may be used. The delimiter should be
2

We shall also use the term CSV for such files, although they are not comma-separated.

206

a normal character (that is, it should be visible and not a so-called control
character). However, a blank or a tab (but not a double-quote) character can
be used as delimiter.
If necessary (so that the resulting file can be loaded by h csv parse data set),
fields will be quoted.

207

208

Chapter 15

Display Information
The HUGIN API was developed partly to satisfy the needs of the HUGIN
GUI application. This application can present an arbitrary belief network or
LIMID model. To do this, it was necessary to associate a certain amount of
graphical information with each node of the network. The functions to
support this are hereby provided for the benefit of the general API user.
Please note that not all items of graphical information have a special interface (such as the one provided for the label of a node see Section 15.1
below). Many more items of graphical information have been added using
the attribute interface described in Section 2.9.2. To find the names of these
extra attributes, take a look at the NET files generated by the HUGIN GUI
application.

15.1

The label of a node

In addition to the name (the syntax of which is restricted), a node can be

assigned an arbitrary string, called the label.
x

h status t h node set label (h node t node, h string t label)

Make a copy of label and assign it as the label of node. There are no restrictions on the contents of the label.
Note that a copy of label is stored inside the node structure, not label itself.
OOBN: If node is an output clone, then the label is not saved if the class is
saved as a NET file (because the NET file format doesnt support that).

h string t h node get label (h node t node)

Returns the label of node. If no label has been associated with node, the
empty string is returned. On error, NULL is returned.
Note that the string returned is the one stored in the node structure. Do not
free it yourself.
209

15.2

The position of a node

In order to display a network graphically, the HUGIN GUI application associates with each node a position in a two-dimensional coordinate system.
The coordinates used by HUGIN are integral values; their type is h coordinate t.
x

h status t h node set position

(h node t node, h coordinate t x, h coordinate t y)
Set the position of node to (x, y).
OOBN: If node is an output clone, then the position is not saved if the class
is saved as a NET file (because the NET file format doesnt support that).

h status t h node get position

(h node t node, h coordinate t x, h coordinate t y)
Retrieve the position (x- and y-coordinates) of node. On error, the values
of x and y are indeterminate.

15.3

The size of a node

As part of a belief network/LIMID specification, the dimensions (width and

height) of a node in a graphical representation of the network can be given.
These parameters apply to all nodes of a domain or a class and are needed
in applications that display the layout of the network in a graphical manner.
An application can modify and inspect these parameters using the functions
described below.
x

h status t h domain set node size

(h domain t domain, size t width, size t height)
Set the width and height dimensions of nodes of domain to width and height,
respectively.

h status t h domain get node size

(h domain t domain, size t width, size t height)
Retrieve the dimensions of nodes of domain. If an error occurs, the values
of variables pointed to by width and height will be indeterminate.
Example 15.1 In an application using a graphical display of a network, a node
could be drawn using the following function.
void draw_node (h_node_t n)
{

210

h_domain_t d = h_node_get_domain (n);

size_t w, h;
h_coordinate_t x, y;
h_domain_get_node_size (d, &w, &h);
h_node_get_position (n, &x, &y);
draw_rectangle (x, y, w, h);
}
Here, draw rectangle is an application-defined function, or maybe a function defined in a graphics library, e.g., XDrawRect if you are using the X Window System.

In a similar way, the width and height dimensions of nodes belonging to

classes in object-oriented models can be accessed.
x

h status t h class set node size

(h class t class, size t width, size t height)
Set the width and height dimensions of nodes of class to width and height,
respectively.

h status t h class get node size

(h class t class, size t width, size t height)
Retrieve the dimensions of nodes of class. If an error occurs, the values of
variables pointed to by width and height will be indeterminate.

211

212

Appendix A

Belief networks with

Conditional Gaussian variables
Beginning with Version 3, the HUGIN API can handle networks with both
discrete and continuous random variables. The continuous random variables must have a Gaussian (also known as a normal) distribution conditional on the values of the parents.
The distribution for a continuous variable Y with discrete parents I and continuous parents Z is a (one-dimensional) Gaussian distribution conditional
on the values of the parents:
P(Y |I = i, Z = z) = N((i) + (i)T z, (i))
Note that the mean depends linearly on the continuous parent variables
and that the variance does not depend on the continuous parent variables.
However, both the linear function and the variance are allowed to depend on
the discrete parent variables. These restrictions ensure that exact inference
is possible.
Discrete variables cannot have continuous parents.
Example A.1 Figure A.1 shows a belief network model for a waste incinerator:
The emissions [of dust and heavy metals] from a waste incinerator
differ because of compositional differences in incoming waste [W].
Another important factor is the waste burning regimen [B], which
can be monitored by measuring the concentration of CO2 in the emissions [C]. The filter efficiency [E] depends on the technical state [F]
of the electrofilter and on the amount and composition of waste [W].
The emission of heavy metals [Mo ] depends on both the concentration
of metals [Mi ] in the incoming waste and the emission of dust particulates [D] in general. The emission of dust [D] is monitored through
measuring the penetrability of light [L]. [20]

213

Mi

Mo

Figure A.1: The structural aspects of the waste incinerator model described
in Example A.1: B, F, and W are discrete variables, while the remaining
variables are continuous.
The result of inference within a belief network model containing Conditional Gaussian variables is the beliefs (i.e., marginal distributions) of the
individual variables given evidence. For a discrete variable this (as usual)
amounts to a probability distribution over the states of the variable. For a
Conditional Gaussian variable two measures are provided:
(1) the mean and variance of the distribution;
(2) since the distribution is in general not a simple Gaussian distribution,
but a mixture (i.e., a weighted sum) of Gaussians, a list of the parameters (weight, mean, and variance) for each of the Gaussians is
available.
The algorithms necessary for computing these results are described in [23].
Example A.2 From the network shown in Figure A.1 (and given that the discrete
variables B, F, and W are all binary), we see that
the distribution for C can be comprised of up to two Gaussians (one if B is
instantiated);
initially (i.e., with no evidence incorporated), the distribution for E is comprised of up to four Gaussians;
if L is instantiated (and none of B, F, or W is instantiated), then the distribution for E is comprised of up to eight Gaussians.

214

Appendix B

HUGIN API Revision History

This appendix contains lists (in reverse chronological order) of changes and
new features in all releases of the HUGIN API since version 2.

Overview of changes and new features in HUGIN API version 8.3

A new data representation tool, called a data set, has been introduced
to facilitate the use of so-called CSV files (short for comma-separatedvalues files) with the learning algorithms in HUGIN. See Chapter 14.
The Boyen-Koller approximation algorithm can now be used for inference in Dynamic Bayesian Networks within the move-time-window
and the prediction operations.
The HUGIN Web Service API has new widgets for the deployment of
Dynamic Bayesian Networks. See the separate online documentation
for more information.

Overview of changes and new features in HUGIN API version 8.2

The PC algorithm for structure learning can now exploit threads running on multi-core (or multi-processor) systems to speed up the learning process. See Section 12.4.
A state-index operator has been introduced in the table generator to
allow using the index of the state of a discrete node within expressions.
This is mostly useful with interval nodes.
A new version of the HKB format is used for networks containing this
new operator (within an expression of the model of a node). If the operator is not used, the format is the same as that used by the previous
version of the HUGIN API.
215

Overview of changes and new features in HUGIN API version 8.1

A COM interface version of the HUGIN API for use with, e.g., Microsoft
Excel is now provided for the Windows platform.
This interface replaces the ActiveX API, which is now deprecated.
The operations for prediction and moving of the time window are now
supported for networks containing function nodes and CG nodes. See
Section 4.3 and Section 4.4.
OOBN models can now be saved as Hugin Knowledge Base files. See
Section 3.13.

Overview of changes and new features in HUGIN API version 8.0

Dynamic models in the form of time-sliced Dynamic Bayesian Networks
are now supported.
A new temporal clone node type is used for expressing temporal
dependencies.
The concept of a sliding time window, containing a fixed number
of time slices, is introduced. This is where evidence is entered
and inference is performed.
A prediction operation computes beliefs for the system beyond
the time window.
The NET file format has been extended in order to represent temporal clones.
See Chapter 4 and Section 13.4.
The intervals of an interval node can now have zero width. This is convenient when the distribution is partly discrete and partly continuous.
See Section 6.6.
HUGIN API libraries for the Windows platforms are now provided for
Visual Studio 2013 (in addition to Visual Studio 6.0, Visual Studio
.NET 2003, Visual Studio 2005, Visual Studio 2008, Visual Studio
2010, and Visual Studio 2012).

Overview of changes and new features in HUGIN API version 7.8

Explanation facilities are now provided: Functions to compute (and
sort) Bayes factors and normalized likelihoods for all evidence subsets
up to a given maximum size are now provided.
See Section 10.12 for further information.
216

 A function to reorder the states of a labeled node is now provided.

The function ensures that data structures (such as tables) are correctly
updated to match the new state ordering.
See h node reorder states(89) .

Overview of changes and new features in HUGIN API version 7.7

HUGIN API libraries for the Windows platforms are now provided for
Visual Studio 2012 (in addition to Visual Studio 6.0, Visual Studio
.NET 2003, Visual Studio 2005, Visual Studio 2008, and Visual Studio
2010).
A new node type has been introduced: The discrete function node type.
This node type is a combination of the discrete and function node types.
The function nodes implemented in previous versions of the HUGIN
API are now referred to as real-valued function nodes.
See Section 2.1.
A new operator to express aggregate distributions has been introduced.
An aggregate distribution is the distribution of the sum of a random
number of independent identically distributed random variables. See
h operator aggregate(81) .
This operator can only be used in models for discrete function nodes.
A new operator to express a probability computed from the beliefs of
a discrete node has been introduced. See h operator probability(81) .
This operator can only be used in models for function nodes.
More general network topologies can now be specified. It is now possible to have a directed path containing functional links between two
nodes that are not real-valued function nodes. This feature is essential
in order to use the new expression operators.
See Section 2.4.
Two new parameters have been introduced for the total-weight triangulation method. These parameters can be used to produce better
triangulations as well as speed up the triangulation process.
An initial triangulation can be specified: If a good triangulation
is already known, then this can be used to speed up the search
for a better triangulation.
The search for minimal separators can be speeded up by discarding separators larger than a prespecified size. This also makes it
possible to handle larger prime components (without having to
split them into smaller graphs).
See Section 7.3 for further details.
217

 A new version of the HKB format is used for networks containing function nodes. If the network does not contain function nodes, the format
is the same as that used by HUGIN API version 7.5.
The NET language has been extended in order to support discrete
function nodes and the new expression operators. See Chapter 13.

Overview of changes and new features in HUGIN API version 7.6

The structure learning algorithm now supports learning of networks
with continuous (CG) nodes.
Function nodes can now be specified as parents of non-function nodes.
This can be used to provide input to the table generation facility. See
Section 6.9.
But in order to avoid cyclic dependencies in the evaluation of function
nodes, some constraints on the network structure must be imposed.
See Section 2.4.
A new version of the HKB format is used for networks containing nonfunction nodes with function parents. If this is not the case, the format
is the same as that used by HUGIN API version 7.5.
The NET language has been extended to allow function nodes as parents of non-function nodes. See Section 13.5 and Section 13.6.

Overview of changes and new features in HUGIN API version 7.5

A web service version of the HUGIN API is now provided. See Chapter 1 and the separate online documentation for more information.
The performance of table operations involving discrete nodes only has
been improved. This has led to improved performance of inference on
many networks.
Function and utility nodes can now be used as inputs and outputs of
classes in OOBNs. See Section 3.7.

Overview of changes and new features in HUGIN API version 7.4

HUGIN API libraries for the Windows platforms are now provided for
Visual Studio 2010 (in addition to Visual Studio 6.0, Visual Studio
.NET 2003, Visual Studio 2005, and Visual Studio 2008).
218

 A function for identifying the requisite ancestors of a decision node

is now provided. This is useful for identifying the relevant past observations and decisions in traditional influence diagrams (i.e., influence
diagrams obeying the no-forgetting rule).
See h node get requisite ancestors(37) .
It is now possible to convert a decision node to a chance node (or vice
versa) while preserving all other attributes of the node unchanged.
See h node set category(30) .
The compression feature is now supported for LIMIDs. See Section 7.6.
The data type used for table indexes within the data structures of compressed domains is now a 32-bit integer type (this used to be a 16-bit
type). This change allows the construction of much larger compressed
tables (at the cost of a higher memory consumption).
Because of this, HKB files containing compressed domains use a new
revision of the file format (but other HKB files use the same revision
as HUGIN API version 7.3).

Overview of changes and new features in HUGIN API version 7.3

A new node type has been introduced: The function node type. This
node type is used to express calculations to be performed using the
results of inference or simulation as input. However, function nodes
are not involved in the inference process itself, so evidence cannot be
specified for function nodes. The exact rules for evaluation of function
nodes are specified in Section 9.7.
Here is a list of changes related to the introduction of function nodes:
The category of function nodes is h category function, and the
kind is h kind other (this kind is also used for utility and instance
nodes). See Section 2.1.
A real-valued function is associated with each function node.
This function is specified using a model (Chapter 6).
The function associated with a function node must only depend
on the parents of the node. The parents can be of any type of
node (except instance nodes). See Section 2.4.
The value of a function node based on the results of inference is
requested by h node get value(128) .
The value of a function node based on the results of simulation
is requested by h node get sampled value(145) .
The h domain new node(30) , h node clone(31) , and h node delete(31)
functions do not uncompile if a function node is involved.
219

 The h node add parent(33) , h node remove parent(34) , and h node

switch parent(34) functions do not uncompile if the child node is
a function node.
The HKB format has been updated to support networks with function nodes (but the old format is still used for networks without
function nodes).
The NET language has been extended: The new function keyword denotes the definition of a function node, and hpotentiali
specifications are now also used to specify links and models for
function nodes. See Chapter 13.
The results of simulation are now invalidated by (implicit as well as
explicit) uncompile operations.
The new h node get sampled utility(146) function requests the sampled utility of a utility node.

Overview of changes and new features in HUGIN API version 7.2

A .NET Compact Framework version of the HUGIN API is now available, providing support for PDAs.
An algorithm for finding the requisite parents of a decision node in
a LIMID has been implemented. See Section 2.4.1.
A Monte Carlo algorithm for finding the most probable configurations
of a set of nodes has been implemented. See Section 10.11.
A new triangulation method has been implemented: This method triangulates each prime component using all of the elimination based
triangulation heuristics and uses the best result. See Section 7.3.
This triangulation method is now used by h domain compile(103) when
compiling untriangulated domains.
As the elimination based triangulation heuristics may produce nonminimal triangulations, an extra pass that removes redundant fill-in
edges has been added to these heuristics.

Overview of changes and new features in HUGIN API version 7.1

The EM algorithm has been extended to learn CG distributions for
continuous nodes.
The EM algorithm now permits learning to be enabled/disabled for
specific parent configurations: A nonnegative experience count enables learning, while a negative experience count disables learning.
220

 The EM algorithm now checks that the equilibrium is sum with no evidence incorporated. As a sanity check, it also verifies that learning is
enabled (that is, there must exist at least one node with a nonnegative
experience count), and that case data has been specified.
Because the EM algorithm is controlled by experience tables, continuous nodes may now also be given experience tables.
The HKB format has been updated (in order to support parameter
learning for continuous nodes).
The model scoring functions (Section 12.3) now check that the junction tree potentials are up-to-date with respect to the node tables and
their models (if any).

Overview of changes and new features in HUGIN API version 7.0

HUGIN API libraries for the Windows platforms are now provided for
Visual Studio 2008 (in addition to Visual Studio 6.0, Visual Studio
.NET 2003, and Visual Studio 2005).
The h domain compile(103) function now initializes the junction trees
using the available evidence (if any).
The h class create domain(58) function now assigns unique meaningful names to the nodes of the created domains. This is accomplished
by allowing dotted names for nodes in domains (but not for nodes
in classes).
This change affects the lexical syntax of node names in files (NET,
data, case, and node list files) and in strings containing expressions.
Sensitivity analysis: It is now possible to compute sensitivity data for
multiple output probabilities simultaneously. This allows for easy solution of constraints on the output probabilities.
Limited memory influence diagrams (LIMIDs) [24] replace influence
diagrams as the tool for modeling decision problems.
This has several implications:
There is no no-forgetting assumption in a LIMID: All information links must be explicitly represented in the network.
A total order of the decisions is no longer required.
Evidence specified for a LIMID must satisfy the free will condition: A chance node cant be observed before all decisions in the
ancestral set of the chance node have been made.
Moreover, only simple instantiations are now permitted as evidence for decision nodes.
See Section 9.1.3 and h error invalid evidence(137) .
221

 Computing optimal decisions is now a separate operation (that

is, they are no longer computed as part of inference). See Section 10.3.
Decision nodes now have tables. These tables represent decision
policies and can be specified in the same way (including the use
of models) as conditional probability and utility tables. However,
they are usually computed by h domain update policies(138) .
In LIMIDs, there are no constraints on elimination orders used
for triangulation.
The overall expected utility of the decision problem is provided
by the new h domain get expected utility(127) function.
h node get belief (125) and h node get expected utility(127) now accept (discrete) chance as well as decision nodes as arguments,
and h node get expected utility also accepts utility nodes.
The usage conditions and the treatment of information links has
changed for d-separation analyses in LIMIDs. See Section 9.4 for
further details.
The nodes argument of h domain get marginal(126) may now contain decision as well as chance nodes. Also, decision nodes are
no longer required to be instantiated and propagated.
The h domain simulate(145) operation no longer requires decision
nodes to be instantiated and (for compiled domains) propagated.
The EM algorithm and the h domain get log likelihood(172) , h domain get AIC(172) , and h domain get BIC(172) functions no longer
require decision nodes to be instantiated in all cases of the data
set. Instead, the evidence in a given case is only required to be
valid.
Sensitivity analysis: Sensitivity functions where the input parameter is associated with a decision node can now be computed.
Also, decision nodes are no longer required to be instantiated
and propagated.
The arguments of h node get entropy(148) and h node get mutual
information(148) may now be decision as well as chance nodes.
Also, decision nodes are no longer required to be instantiated
and propagated.
The h domain save to memory(142) operation is disabled for LIMIDs.
Calculation of conflict of evidence is not supported for LIMIDs.
The NET format has been extended in order to represent policies.
The HKB format has changed.
222

Overview of changes and new features in HUGIN API version 6.7

Sensitivity analysis: Functions to aid in the process of identifying the
most influential (conditional probability) parameters of a belief network model and analyzing their effects on the output probabilities
of the model are now provided.
See Section 10.10.
New statistical distributions: LogNormal, Triangular, and PERT.
See Section 6.7.1.
An optional location parameter has been added to the Gamma, Exponential, and Weibull distributions.
See Section 6.7.1.
Truncated continuous distributions can be expressed using a new truncation operator: It is possible to specify double as well as single truncation (single truncation is either left or right truncation).
See h operator truncate(80) .
It is now possible to combine expressions of different types in conditional expressions. This permits specification of relationships that are
sometimes probabilistic and sometimes deterministic.
See h operator if (82) .
The HKB format has changed (because of the new expression operators).
The performance of h domain get marginal(126) has been improved.
The performance of inference has been improved for the case where a
memory backup is not available.
A memory backup can no longer be created when evidence has been
propagated.

Overview of changes and new features in HUGIN API version 6.6

A new .NET API is now available for the Windows platforms. This API
targets the .NET 2.0 framework.
HUGIN API libraries for the Windows platforms are now provided for
Visual Studio 2005 (in addition to Visual Studio 6.0 and Visual Studio
.NET 2003).
The HUGIN APIs (except the Visual Basic API) are now available as
64-bit versions on all platforms except Mac OS X.
On the Windows platforms, the 64-bit libraries are only provided for
Visual Studio 2005.
223

 A new naming scheme has been introduced for the HUGIN API libraries on the Windows platforms: The libraries are now uniquely
named, making it possible to have all DLLs in the search path simultaneously.
d-separation analysis is now used to improve the performance of inference. This is particularly useful for incremental propagation of evidence in large networks.
The performance of the total-weight triangulation method has been
greatly improved.
The triangulation functions now construct junction trees, but do not
allocate storage for the data arrays of the clique and separator tables.
This permits the application to see the junction trees before attempting
the final (and most expensive) part of the compilation process.
It is now possible to query the size of a junction tree (even before
storage is allocated for the junction tree tables). See h jt get total
size(117) and h jt get total cg size(117) .
A function h domain is triangulated(108) for testing whether a domain is triangulated is now provided.
The HUGIN KB file format has changed (in order to handle HKB files
produced by 64-bit versions of the HUGIN API, among other things).
There are a few user-visible changes: If a compiled (but not compressed) domain is saved as an HKB file, it will only be triangulated
when loaded. A compilation is required before inference can be performed (see Section 2.10). Compressed domains are still loaded as
compressed (which implies compiled), but a propagation is required
before beliefs can be retrieved.
Functions for converting between table indexes and state configurations are now provided: h table get index from configuration(71) and
h table get configuration from index(71) .
A function to retrieve the CG size of a table is provided: h table get
cg size(74) .

Overview of changes and new features in HUGIN API version 6.5

Domains and nodes can now be constructed by cloning existing objects see h domain clone(29) and h node clone(31) .
Default state labels for boolean nodes are now false and true
see h node get state label(88) .1
1
This was done in order to ensure compatibility between the different versions of the
HUGIN API.

224

 A function for translating a state label to a state index is provided

see h node get state index from label(88) . Also, a function for translating a state value to a state index is provided see h node get state
index from value(90) .
Functions for determining independence properties (also known as dseparation properties) are provided see h domain get d connected
nodes(124) and h domain get d separated nodes(124)
The number of cases that can be handled using a given amount of
memory has been doubled compared to previous releases. This has
been made possible by limiting the number of different states that can
be represented in case data to 32768 per node see h node set case
state(166) .
A function for entering a case as evidence is provided see h domain
enter case(168) .
h domain get log likelihood(172) now computes the log-likelihood with
respect to the current parameter values. It used to return the loglikelihood computed during the final iteration of the EM algorithm
(implying that the log-likelihood was computed with respect to slightly
outdated parameter values). The function no longer requires invoking
the EM algorithm before use.
Functions for computing the AIC and BIC scores are provided see
h domain get AIC(172) and h domain get BIC(172) . AIC and BIC are
scores of comparing model quality taking model complexity into account.
The EM algorithm now reports the AIC and BIC scores (in addition to
the log-likelihood score) to the log file after completion of parameter
estimation.
The formats of case files and data files have been extended: If the
contents of a state label form a valid name, the quotes can be omitted.
See Section 9.9 and Section 12.2.

Overview of changes and new features in HUGIN API version 6.4

The performance of inference (including the EM algorithm) has been
improved.
Functions for computing entropy and mutual information have been
added: h node get entropy(148) and h node get mutual information(148) .
These functions are useful for value of information analysis. See Section 10.9 for further details.
New function: h domain get log likelihood(172) .
225

 h domain learn tables(176) and h domain learn class tables(179) now report the log-likelihood to the log-file (if it is non-NULL) after each iteration of the EM algorithm.
Direct access to the pseudorandom number generator implemented in
Hugin is now provided through the functions h domain get uniform
deviate(146) and h domain get normal deviate(146) .
HUGIN API libraries for Windows platforms are now provided for Visual Studio .NET 2003 (in addition to Visual Studio 6.0).

Overview of changes and new features in HUGIN API version 6.3

A function for replacing the class of an instance node with another
compatible class (i.e., the interface of the new class must be a superset
of the interface of the class being replaced), while preserving all input
and output relations associated with the instance node. This is useful
when, for example, a new version of an instantiated class becomes
available. See h node substitute class(57) .
A function for replacing a parent of a node with another compatible
node, while preserving the validity of existing tables and model associated with the child node. See h node switch parent(34) .
The C++ HUGIN API is now available as both a single-precision and a
double-precision library.

Overview of changes and new features in HUGIN API version 6.2

HUGIN KB files are now automatically compressed using the Zlib library (www.zlib.net). This change implies that the developer (i.e.,
the user of the HUGIN API) must explicitly link to the Zlib library, if
the application makes use of HKB files. See Section 1.2.
HUGIN KB files can now be protected by a password. The following
new functions supersede old functions: h domain save as kb(49) and
h kb load domain(49) .
The EM learning algorithm can now be applied to object-oriented
models. See the h domain learn class tables(179) function.
Functions to save and load case data (as used by the learning algorithms Section 12.1) have been added to the API. See Section 12.2.
Functions to parse a list of node names stored in a text file are now
provided. Such functions are useful for handling, e.g., collections of
elimination orders for triangulations. See h domain parse nodes(108)
and h class parse nodes(108) .
226

 The HUGIN API Reference Manual is now provided as a hyperlinked

PDF file.

Overview of changes and new features in HUGIN API version 6.1

The HUGIN API is now thread-safe. See Section 1.8 for further details.
Functions to save and load cases have been added to the API. See
Section 9.9.
The heuristic used in the total-weight triangulation method for large
networks has been improved.

Overview of changes and new features in HUGIN API version 6

Object-oriented models for belief networks and influence diagrams
can now be constructed using HUGIN API functions (see Chapter 3).
Also, NET language support for object-oriented models (including generation and parsing of NET specifications) is available (see Chapter 13).
Support for API functions prior to Version 2 of the HUGIN API has
been dropped.
Loading of HKB files produced by API versions prior to version 5.0 has
been dropped. If you have an old release, please save your domains
using the NET format before upgrading.
Some functions have been superseded by better ones: h domain write
net has been replaced by h domain save as net(201) , h net parse has
been replaced by h net parse domain(198) , and h string to expression
has been replaced by h string parse expression(86) . However, the old
functions still exist in the libraries, but the functions should not be
used in new applications.

Overview of changes and new features in HUGIN API version 5.4

A new triangulation method has been implemented. This method
makes it possible to find a (minimal) triangulation with minimum
sum of clique weights. For some large networks, this method has improved time and space complexity of inference by an order of magnitude (sometimes even more), compared to the heuristic methods
provided by earlier versions of the HUGIN API.
See Section 7.3 for more information.
The computations used in the inference process have been reorganized
to make better use of the caches in modern CPUs. The result is faster
inference.
227

Overview of changes and new features in HUGIN API version 5.3

The structure learning algorithm now takes advantage of domain knowledge in order to constrain the set of possible networks. Such knowledge can be knowledge of the direction of an edge, the presence or
absence of an edge, or both. See Section 12.5.
A new operator (Distribution) for specifying arbitrary finite discrete
distributions has been introduced. This operator is only permitted for
discrete variables (i.e., not interval variables).
The discrete distributions (Binomial, Poisson, Negative Binomial, and
Geometric) now also work for interval nodes.
New functions for the table generator: The floor and ceil functions
round real numbers to integers, the abs function computes the absolute value of a number, and the mod (modulo) function computes
the remainder of a division.
The HUGIN KB file format has changed (again), but version 5.3 of the
HUGIN API will load HKB files produced by versions 3 or later (up to
version 5.3). But note that support for older formats may be dropped
in future versions of the HUGIN API.

Overview of changes and new features in HUGIN API version 5.2

An algorithm for learning the structure of a belief network given a set
of cases has been implemented. See Section 12.4.
Simulation in uncompiled domains (Section 10.8) is now permitted
when the set of nodes with evidence form an ancestral set of instantiated nodes (i.e., no likelihood evidence is present, and if a chance
node is instantiated, so are all of its parents). Decision nodes must, of
course, be instantiated.
If a domain is saved (as a HUGIN KB file) in compiled form, h kb load
domain(49) attempts to load it in that form as well. As the contents of
the junction tree tables are not stored in the HKB file, the inference
engine must be initialized from the user-specified tables and models.
This can fail for various reasons (e.g., the tables and/or models contain invalid data). In this case, instead of refusing to load the domain,
h kb load domain instead returns the domain in uncompiled form.
The log2 , log10 , sin, cos, tan, sinh, cosh, and tanh functions and the
Negative Binomial distribution have been added to the table generation facility.
228

 The HUGIN KB file format has changed (again), but version 5.2 of the
HUGIN API will load HKB files produced by versions 3 or later (up to
version 5.2). But note that support for older formats may be dropped
in future versions of the HUGIN API.

Overview of changes and new features in HUGIN API version 5.1

The simulation procedure has been extended to handle networks with
continuous variables. Also, a method for simulation in uncompiled
domains has been added. See Section 10.8.
HUGIN will now only generate tables from a model when (the inference engine thinks) the generated table will differ from the most
recently generated table. Such tests are now performed by the compilation, propagation, and reset-inference-engine operations (in previous versions of the HUGIN API, the compilation operation always
generated all tables, and the propagation and reset-inference-engine
operations never generated any tables).
Also, tables can now be [re]generated individually on demand.
See Section 6.8 for more information.
The number of values to use per (bounded) interval of a (parent) interval node can now be specified on a per-model basis. This provides
a way to trade accuracy for computation speed. See Section 6.9.
Iterators for the attributes of nodes and domains are now provided.
See Section 2.9.2.
The HUGIN KB file format (the .hkb files) has changed. This was
done in order to accommodate the above mentioned features.
The HUGIN API version 5.1 will load HUGIN KB files produced by
versions 3 or later (up to version 5.1). But note that support for older
formats may be dropped in future versions of the HUGIN API.
The NET language has been extended with a model attribute for specifying the number of values to use per (bounded) interval of a (parent)
interval node.
Also, if both a specification using the model attributes and a specification using the data attribute are provided, then the specification in
the data attribute is used. Previous versions of the HUGIN API used
the model in such cases.
See Section 13.6.2 for more information.
229

Overview of changes and new features in HUGIN API version 5

A batch learning method (based on the EM algorithm) has been implemented. Given a set of cases and optional expert-supplied priors, it
finds2 the best unrestricted model matching the data and the priors.
The sequential learning method (also known as adaptation) has been
reimplemented and given a new API interface. (HUGIN API version
1.2 provided the first implementation of sequential learning.)
The HUGIN KB file format (the .hkb files) has changed. This was
done in order to accommodate adaptation information and evidence.
Also, junction tree tables (for compiled domains) are not stored in the
HUGIN KB file anymore.
The HUGIN API version 5 will load HUGIN KB files produced by HUGIN
API versions 3 or later (up to version 5).
The NET language has been extended in order to accommodate adaptation information.
A single-precision version of the HUGIN API is now able to load a
HUGIN KB file created by a double-precision version of the HUGIN
API and vice versa.

Overview of changes and new features in HUGIN API version 4.2

The traditional function-oriented version of the HUGIN API has been
supplemented by object-oriented versions for the Java and C++ language environments.
The most time-consuming operations performed during inference have
been made threaded. This makes it possible to speed up inference by
having individual threads execute in parallel on multi-processor systems.
The class of belief networks with CG nodes that can be handled by
HUGIN has been extended. A limitation of the old algorithm has been
removed by the introduction of the recursive combination operation
(see [23] for details).
Evidence entered to a domain is no longer removed when an (explicit
or implicit) uncompile operation is performed. Also, evidence can
be entered (and retracted) when the domain is not compiled. These
2

The EM algorithm is an iterative method that searches for a maximum of a function.

There is, however, no guarantee that the maximum is global. It might be a local maximum
or even a saddle point.

230

changes affect all functions that enter, retract, or query (entered) evidence, as well as h domain uncompile(110) and the functions that perform implicit uncompile operations with the exception of h node
set number of states(38) which still removes the entered evidence.

Overview of changes and new features in HUGIN API version 4.1

An arc-reversal operation is provided: This permits the user to reverse the direction of an edge between two chance nodes of the same
kind, while at the same time preserving the joint probability distribution of the belief network or influence diagram.
A Noisy OR distribution has been added to the table generation facility (Chapter 6).
Support for C compilers that dont conform to the ISO Standard C
definition has been dropped.

Overview of new features in HUGIN API version 4

Version 4 of the HUGIN API makes it possible to generate conditional
probability and utility potentials based on mathematical descriptions
of the relationships between nodes and their parents. The language
provided for such descriptions permits both deterministic and probabilistic relationships to be expressed.
This facility is implemented as a front end to the HUGIN inference engine: The above mentioned descriptions only apply to discrete nodes,
implying that continuous distributions (such as the gamma distribution) are discretized. Thus, inference with such distributions are only
approximate. The only continuous distributions for which the HUGIN
API provides exact inference are the CG distributions.
See Chapter 6 for further details.
The table for a node is no longer deleted when a parent is removed or
the number of states is changed (either for the node itself or for some
parent). Instead, the table is resized (and the contents updated).
This change affects the following functions: h node remove parent(34) ,
h node set number of states(38) , and h node delete(31) (since deletion
of a node implies removing it as a parent of its children).

Overview of new features in HUGIN API version 3

Version 3 of the HUGIN API introduces belief networks with Conditional Gaussian (CG) nodes. These represent variables with a Gaussian (also known as a normal) distribution conditional on the values
231

of their parents. The inference is exact (i.e., no discretization is performed).

It is no longer required to keep a copy of the initial distribution stored
in a disk file or in memory in order to initialize the inference engine.
Instead, the initial distribution can be computed (when needed) from
the conditional probability and utility tables.
It is now possible for the user to associate attributes (key/value pairs)
with nodes and domains. The advantage over the traditional user data
(as known from previous versions of the HUGIN API) is that these
attributes are saved with the domain in both the NET and the HUGIN
KB formats.
It is no longer necessary to recompile a domain when some conditional
probability or utility potential has changed. When HUGIN notices that
some potential has changed, the updated potential will be taken into
account in subsequent propagations.
It is now possible to reorganize the layout of conditional probability
and utility potentials (Section 5.5).
The HUGIN API is now provided in two versions: A (standard) version using single-precision floating-point arithmetic and a version using double-precision floating-point arithmetic. The double-precision
version may prove useful in computations with continuous random
variables (at the cost of a larger space requirement).

Overview of new features in HUGIN API version 2

Version 2 of the HUGIN API introduces influence diagrams. An influence diagram is a belief network augmented with decisions and utilities. Edges in the influence diagram into a random variable represents
probabilistic dependencies while edges into a decision variable represents availability of information at the time the decision is taken.
Assuming a total order of the decisions, an optimal decision policy
using maximization of expected utility for choosing between decision
alternatives can be computed.
Version 2 of the HUGIN API allows specification of and inference and
decision making with influence diagrams. This version will also take
advantage of the case where the overall utility is a sum of a set of local
utilities.
New propagation methods: (1) In addition to the well-known sumpropagation method, a max-propagation method that identifies the
most probable configuration of all variables and computes its probability is introduced, and (2) a new way to incorporate evidence, known
232

as fast-retraction, permits the computation, for each variable, of the

conditional probability of that variable given evidence on the remaining variables (useful for identifying suspicious findings). Thus, four
different ways of propagating evidence are now available.
Models with undirected edges, so-called chain graph models, are now
permitted. This extends the class of models so that automatically generated models are more easily used with HUGIN (in automatically generated models, the direction of an association is only rarely identified).
Chain graph models are currently only available via NET specifications
(Chapter 13).
Extraction of the joint probability distribution for a group of variables,
even when the group is not a subset of any clique, is now possible.
Version 2 of the HUGIN API allows construction and editing of belief
networks and influence diagrams.
Analysis of data conflicts, previously only available within the HUGIN
GUI application, is now also available via the API.
Simulation: Given evidence, a configuration for all variables can be
sampled according to the distribution determined by the evidence.
The interface of the API has undergone a major clean-up and redesign.
The naming has been made more consistent: A common prefix h is
introduced, all functions operating on the same type of object has a
common prefix (e.g., all functions with a node as primary argument
shares the common prefix h node )
The concept of a current or selected domain has been removed. The
domain to be operated upon is now an explicit argument.
Backwards compatibility: Application programs built using the documented functions and types of previous versions of the HUGIN API can
still be compiled and should work as expected, although use of these
older functions and types in new applications is strongly discouraged.

233

234

Bibliography
[1] S. K. Andersen, K. G. Olesen, F. V. Jensen, and F. Jensen. HUGIN
a shell for building Bayesian belief universes for expert systems. In
Proceedings of the Eleventh International Joint Conference on Artificial
Intelligence, pages 10801085, Detroit, Michigan, Aug. 2025, 1989.
Reprinted in [33].
[2] A. Berry, J.-P. Bordat, and O. Cogis. Generating all the minimal separators of a graph. International Journal of Foundations of Computer
Science, 11(3):397403, Sept. 2000.
[3] V. Bouchitte and I. Todinca. Treewidth and minimum fill-in: Grouping
the minimal separators. SIAM Journal on Computing, 31(1):212232,
July 2001.
[4] X. Boyen and D. Koller. Tractable inference for complex stochastic processes. In G. F. Cooper and S. Moral, editors, Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence, pages 3342,
Madison, Wisconsin, July 2426, 1998. Morgan Kaufmann, San Francisco, California.
[5] H. Chan and A. Darwiche. When do numbers really matter? Journal
of Artificial Intelligence Research, 17:265287, 2002.
[6] V. M. H. Coupe and L. C. van der Gaag. Properties of sensitivity analysis of Bayesian belief networks. Annals of Mathematics and Artificial
Intelligence, 36(4):323356, Dec. 2002.
[7] R. G. Cowell and A. P. Dawid. Fast retraction of evidence in a probabilistic expert system. Statistics and Computing, 2(1):3740, Mar.
1992.
[8] R. G. Cowell, A. P. Dawid, S. L. Lauritzen, and D. J. Spiegelhalter.
Probabilistic Networks and Expert Systems. Statistics for Engineering
and Information Science. Springer-Verlag, New York, 1999.
[9] A. Darwiche. Modeling and Reasoning with Bayesian Networks. Cambridge University Press, Cambridge, UK, 2009.
235

[10] A. P. Dawid. Applications of a general propagation algorithm for probabilistic expert systems. Statistics and Computing, 2(1):2536, Mar.
1992.
[11] F. Jensen and S. K. Andersen. Approximations in Bayesian belief universes for knowledge-based systems. In Proceedings of the Sixth Conference on Uncertainty in Artificial Intelligence, pages 162169, Cambridge, Massachusetts, July 2729, 1990.
[12] F. V. Jensen, B. Chamberlain, T. Nordahl, and F. Jensen. Analysis in
HUGIN of data conflict. In P. P. Bonissone, M. Henrion, L. N. Kanal, and
J. F. Lemmer, editors, Uncertainty in Artificial Intelligence, volume 6,
pages 519528. Elsevier Science Publishers, Amsterdam, The Netherlands, 1991.
[13] F. V. Jensen, S. L. Lauritzen, and K. G. Olesen. Bayesian updating in
causal probabilistic networks by local computations. Computational
Statistics Quarterly, 5(4):269282, 1990.
[14] F. V. Jensen and T. D. Nielsen. Bayesian Networks and Decision Graphs.
Information Science and Statistics. Springer-Verlag, New York, second
edition, 2007.
[15] F. V. Jensen, K. G. Olesen, and S. K. Andersen. An algebra of Bayesian
belief universes for knowledge-based systems. Networks, 20(5):637
659, Aug. 1990. Special Issue on Influence Diagrams.
[16] U. Kjrulff. Triangulation of graphs algorithms giving small total
state space. Research Report R-90-09, Department of Mathematics
and Computer Science, Aalborg University, Denmark, Mar. 1990.
[17] U. Kjrulff. dHugin: a computational system for dynamic time-sliced
Bayesian networks. International Journal of Forecasting, 11(1):89
111, Mar. 1995. Special issue on Probability Forecasting.
[18] U. B. Kjrulff and A. L. Madsen. Bayesian Networks and Influence Diagrams: A Guide to Construction and Analysis. Information Science and
Statistics. Springer-Verlag, New York, second edition, 2013.
[19] D. Koller and N. Friedman. Probabilistic Graphical Models: Principles
and Techniques. Adaptive Computation and Machine Learning. MIT
Press, Cambridge, Massachusetts, 2009.
[20] S. L. Lauritzen. Propagation of probabilities, means, and variances in
mixed graphical association models. Journal of the American Statistical
Association (Theory and Methods), 87(420):10981108, Dec. 1992.
236

[21] S. L. Lauritzen. The EM algorithm for graphical association models with missing data. Computational Statistics & Data Analysis,
19(2):191201, Feb. 1995.
[22] S. L. Lauritzen, A. P. Dawid, B. N. Larsen, and H.-G. Leimer. Independence properties of directed Markov fields. Networks, 20(5):491505,
Aug. 1990. Special Issue on Influence Diagrams.
[23] S. L. Lauritzen and F. Jensen. Stable local computation with conditional Gaussian distributions. Statistics and Computing, 11(2):191
203, Apr. 2001.
[24] S. L. Lauritzen and D. Nilsson. Representing and solving decision
problems with limited information. Management Science, 47(9):1235
1251, Sept. 2001.
[25] S. L. Lauritzen and D. J. Spiegelhalter. Local computations with probabilities on graphical structures and their application to expert systems. Journal of the Royal Statistical Society, Series B (Methodological),
50(2):157224, 1988. Reprinted in [33].
[26] A. L. Madsen, F. Jensen, M. Karlsen, and N. Sndberg-Jeppesen.
Bayesian networks with function nodes. In L. C. van der Gaag and
A. J. Feelders, editors, PGM 2014, volume 8754 of LNAI, pages 286
301, Utrecht, The Netherlands, Sept. 1719, 2014. Springer-Verlag.
[27] A. L. Madsen, F. Jensen, U. B. Kjrulff, and M. Lang. The HUGIN tool
for probabilistic graphical models. International Journal on Artificial
Intelligence Tools, 14(3):507543, June 2005.
[28] A. L. Madsen, M. Lang, U. B. Kjrulff, and F. Jensen. The HUGIN tool
for learning Bayesian networks. In T. D. Nielsen and N. L. Zhang, editors, ECSQARU 2003, volume 2711 of LNAI, pages 594605, Aalborg,
Denmark, July 25, 2003. Springer-Verlag.
[29] K. G. Olesen, S. L. Lauritzen, and F. V. Jensen. aHUGIN: A system
creating adaptive causal probabilistic networks. In D. Dubois, M. P.
Wellman, B. DAmbrosio, and P. Smets, editors, Proceedings of the
Eighth Conference on Uncertainty in Artificial Intelligence, pages 223
229, Stanford, California, July 1719, 1992. Morgan Kaufmann, San
Mateo, California.
[30] K. G. Olesen and A. L. Madsen. Maximal prime subgraph decomposition of Bayesian networks. IEEE Transactions on Systems, Man, and
Cybernetics, Part B: Cybernetics, 32(1):2131, Feb. 2002.
[31] J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference. Morgan Kaufmann, San Mateo, California, 1988.
237

[32] J. Pearl. Causality: Models, Reasoning, and Inference. Cambridge University Press, Cambridge, UK, 2000.
[33] G. Shafer and J. Pearl, editors. Readings in Uncertain Reasoning. Morgan Kaufmann, San Mateo, California, 1990.
[34] K. Shoikhet and D. Geiger. A practical algorithm for finding optimal
triangulations. In Proceedings of the Fourteenth National Conference on
Artificial Intelligence, pages 185190, Providence, Rhode Island, July
2731, 1997. AAAI Press, Menlo Park, California.
[35] D. J. Spiegelhalter and S. L. Lauritzen. Sequential updating of conditional probabilities on directed graphical structures. Networks,
20(5):579605, Aug. 1990. Special Issue on Influence Diagrams.
[36] P. Spirtes, C. Glymour, and R. Scheines. Causation, Prediction, and
Search. Adaptive Computation and Machine Learning. MIT Press,
Cambridge, Massachusetts, second edition, 2000.
[37] D. Vose. Risk Analysis: A Quantitative Guide. Wiley, Chichester, UK,
second edition, 2000.

238

Index
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

attribute get key, 47

attribute get next, 47
attribute get value, 47
attribute t, 46
boolean make expression, 79
category chance, 28
category decision, 28
category error, 28
category function, 28
category instance, 28, 55
category utility, 28
cc delete, 52
cc get class by name, 53
cc get members, 52
cc new class, 52
cc save as kb, 62
cc save as net, 201
class collection t, 51
class create dbn domain, 65
class create domain, 58
class delete, 52
class generate tables, 96
class get attribute, 62
class get class collection, 52
class get file name, 201
class get first attribute, 62
class get first node, 60
class get inputs, 54
class get instances, 55
class get name, 53
class get node by name, 53
class get node size, 211
class get outputs, 54
class get user data, 61
class new instance, 55
class new node, 53

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
239

class parse nodes, 108

class save as net, 201
class set attribute, 61
class set log file, 97
class set name, 52
class set node size, 211
class set user data, 61
class t, 51
clique get junction tree, 116
clique get members, 117
clique get neighbors, 118
clique t, 116
constraint backward edge forbidden, 175
constraint backward edge required,
175
constraint edge forbidden, 175
constraint edge required, 175
constraint error, 175
constraint forward edge forbidden,
175
constraint forward edge required,
175
constraint none, 175
coordinate t, 210
count t, 17
csv parse data set, 206
data set t, 204
domain adapt, 162
domain add cases, 205
domain approximate, 112
domain cg evidence is propagated,
144
domain clone, 29
domain compile, 103
domain compress, 110

h domain compute dbn predictions,

67
h domain compute sensitivity data,
153
h domain delete, 29
h domain enter case, 168
h domain equilibrium is, 143
h domain evidence is propagated, 144
h domain evidence mode is, 143
h domain evidence to propagate, 144
h domain find map configurations,
154
h domain generate tables, 96
h domain get AIC, 172
h domain get approximation constant,
113
h domain get attribute, 46
h domain get BIC, 172
h domain get case count, 167
h domain get concurrency level, 21
h domain get conflict, 139
h domain get d connected nodes, 124
h domain get d separated nodes, 124
h domain get dbn window offset, 66
h domain get elimination order, 108
h domain get expected utility, 127
h domain get explanation, 158
h domain get explanation score, 158
h domain get file name, 201
h domain get first attribute, 46
h domain get first junction tree, 116
h domain get first node, 43
h domain get grain size, 22
h domain get log likelihood, 172
h domain get log likelihood tolerance,
178
h domain get log normalization constant, 140
h domain get map configuration, 155
h domain get marginal, 126
h domain get max number of em iterations, 179
h domain get max number of separators, 106
240

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

domain get max separator size, 107

domain get node by name, 43
domain get node size, 210
domain get normal deviate, 146
domain get normalization constant,
140
domain get number of cases, 166
domain get number of explanations,
158
domain get number of map configurations, 155
domain get probability of map configuration, 156
domain get sensitivity set, 151
domain get sensitivity set by output, 154
domain get significance level, 174
domain get uniform deviate, 146
domain get user data, 45
domain initialize, 142
domain initialize dbn window, 66
domain is compiled, 103
domain is compressed, 111
domain is triangulated, 108
domain learn class tables, 179
domain learn structure, 173
domain learn tables, 176
domain likelihood is propagated,
144
domain move dbn window, 66
domain new case, 166
domain new node, 30
domain parse case, 132
domain parse cases, 169
domain parse nodes, 108
domain propagate, 135
domain reset inference engine, 142
domain retract findings, 122
domain save as kb, 49
domain save as net, 201
domain save case, 132
domain save cases, 170
domain save to memory, 142
domain seed random, 146

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
H
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

domain set attribute, 46

domain set case count, 167
domain set concurrency level, 21
domain set grain size, 21
domain set initial triangulation, 106
domain set log file, 109
domain set log likelihood tolerance,
178
domain set max number of em iterations, 179
domain set max number of separators, 106
domain set max separator size, 107
domain set node size, 210
domain set number of cases, 166
domain set significance level, 174
domain set user data, 45
domain simulate, 145
domain t, 27
domain tables to propagate, 144
domain triangulate, 107
domain triangulate dbn, 65
domain triangulate with order, 107
domain uncompile, 110
domain update policies, 138
DOUBLE, 3, 4, 7, 8, 10, 11
double t, 17
ds delete, 204
ds delete column, 204
ds delete row, 204
ds get column name, 204
ds get data item, 205
ds get number of columns, 204
ds get number of rows, 204
ds move column, 205
ds move row, 205
ds new column, 204
ds new row, 204
ds save, 206
ds set column name, 204
ds set data item, 205
edge constraint t, 174
equilibrium max, 134
equilibrium sum, 134
241

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

equilibrium t, 134
error code, 18
error compressed, 111
error description, 19
error fast retraction, 136
error inconsistency or underflow,
136
error invalid evidence, 137
error io, 21
error name, 19
error no memory, 21
error none, 18
error overflow, 136
error t, 18
error usage, 20
evidence mode t, 135
expression clone, 84
expression delete, 84
expression get boolean, 84
expression get label, 84
expression get node, 84
expression get number, 84
expression get operands, 84
expression get operator, 83
expression is composite, 83
expression t, 78
expression to string, 86
index t, 18
infinity, 90
jt cg evidence is propagated, 144
jt equilibrium is, 143
jt evidence is propagated, 144
jt evidence mode is, 144
jt evidence to propagate, 144
jt get cliques, 116
jt get conflict, 139
jt get next, 116
jt get root, 117
jt get total cg size, 117
jt get total size, 117
jt likelihood is propagated, 144
jt propagate, 137
jt tables to propagate, 145
junction tree t, 116

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

kb load class collection, 62

kb load domain, 49
kind continuous, 29
kind discrete, 29
kind error, 29
kind other, 29
label make expression, 79
location t, 86, 198
make composite expression, 79
mode fast retraction, 135
mode normal, 135
model delete, 87
model get expression, 88
model get nodes, 87
model get number of samples per
interval, 98
model get size, 87
model set expression, 87
model set number of samples per
interval, 98
model t, 86
net parse classes, 199
net parse domain, 198
new class collection, 52
new data set, 204
new domain, 29
node add parent, 33
node add to inputs, 54
node add to outputs, 54
node case is set, 167
node category t, 28
node clone, 31
node compute explanation data, 157
node compute sensitivity data, 150
node create temporal clone, 63
node delete, 31
node enter finding, 122
node enter value, 122
node evidence is entered, 130
node evidence is propagated, 130
node evidence to propagate, 144
node generate table, 95
node get alpha, 42
node get attribute, 46
242

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node
node

get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get
get

belief , 125
beta, 42
case state, 166
case value, 167
category, 30
children, 36
distribution, 126
domain, 30
edge constraint, 175
entered finding, 129
entered value, 129
entropy, 148
expected utility, 127
experience table, 160
fading table, 161
first attribute, 46
gamma, 42
home class, 53
input, 58
instance, 57
instance class, 55
junction tree, 116
kind, 30
label, 209
master, 56
mean, 125
model, 87
mutual information, 148
name, 42
next, 43
number of states, 39
output, 57
parents, 35
position, 210
predicted belief , 67
predicted mean, 68
predicted value, 68
predicted variance, 68
propagated finding, 129
propagated value, 130
requisite ancestors, 37
requisite parents, 36
sampled state, 145
sampled utility, 146

h node get sampled value, 145

h node get sensitivity constants, 150
h node get sensitivity constants by
output, 153
h node get source, 59
h node get state index from label, 88
h node get state index from value, 90
h node get state label, 88
h node get state value, 90
h node get subtype, 78
h node get table, 39
h node get temporal clone, 64
h node get temporal master, 64
h node get user data, 44
h node get value, 128
h node get variance, 125
h node has experience table, 160
h node has fading table, 161
h node kind t, 28
h node likelihood is entered, 130
h node likelihood is propagated, 130
h node make expression, 78
h node new model, 86
h node remove from inputs, 54
h node remove from outputs, 54
h node remove parent, 34
h node reorder states, 89
h node retract findings, 122
h node reverse edge, 35
h node select state, 121
h node set alpha, 41
h node set attribute, 46
h node set beta, 41
h node set case state, 166
h node set case value, 167
h node set category, 30
h node set edge constraint, 175
h node set gamma, 42
h node set input, 57
h node set label, 209
h node set name, 42
h node set number of states, 38
h node set position, 210
h node set state label, 88
243

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

node set state value, 89

node set subtype, 78
node set user data, 44
node substitute class, 57
node subtype t, 78
node switch parent, 34
node t, 27
node touch table, 41
node unset case, 167
node unset input, 58
number make expression, 79
number t, 17
operator abs, 82
operator add, 79
operator aggregate, 81
operator and, 83
operator Beta, 80
operator Binomial, 80
operator boolean, 83
operator ceil, 82
operator cos, 82
operator cosh, 82
operator Distribution, 80
operator divide, 79
operator equals, 79
operator error, 83
operator exp, 82
operator Exponential, 80
operator floor, 82
operator Gamma, 80
operator Geometric, 80
operator greater than, 79
operator greater than or equals, 79
operator if , 82
operator label, 83
operator less than, 79
operator less than or equals, 79
operator log, 82
operator log10, 82
operator log2, 82
operator LogNormal, 80
operator max, 82
operator min, 82
operator mod, 82

h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h
h

operator multiply, 79
operator negate, 79
operator NegativeBinomial, 80
operator node, 83
operator NoisyOR, 80
operator Normal, 80
operator not, 83
operator not equals, 79
operator number, 83
operator or, 83
operator PERT, 80
operator Poisson, 80
operator power, 79
operator probability, 81
operator sin, 82
operator sinh, 82
operator sqrt, 82
operator state index, 80
operator subtract, 79
operator t, 79
operator tan, 82
operator tanh, 82
operator Triangular, 80
operator truncate, 80
operator Uniform, 80
operator Weibull, 80
status t, 17
string parse expression, 86
string t, 17
subtype boolean, 78
subtype error, 78
subtype interval, 78
subtype label, 78
subtype number, 78
table delete, 73
table get cg size, 74
table get configuration from index,
71
table get covariance, 73
table get data, 72
table get index from configuration,
71
table get mean, 73
table get nodes, 72
244

h
h
h
h
h
h
h
h
h
h
h

table get size, 73

table get variance, 73
table reorder nodes, 74
table t, 69
tm best greedy, 105
tm clique size, 104
tm clique weight, 104
tm fill in size, 104
tm fill in weight, 104
tm total weight, 105
triangulation method t, 104

Linux, 2
Mac OS X, 2
Solaris, 2, 2223
UTF-8, 45, 185, 196
Windows, 515, 23
Zlib, 2, 3