Vehicle Circulation and

the Hungarian Method

Martin Grtschel
joint work with
Ralf Borndrfer Andreas Lbel

Celebration Day of the 50th Anniversary

of the Hungarian Method
Budapest, October 31, 2005

Martin Grtschel Institute of Mathematics, Technische Universitt Berlin (TUB)

DFG-Research Center Mathematics for key technologies (MATHEON)
Konrad-Zuse-Zentrum fr Informationstechnik Berlin (ZIB)
[email protected] http://www.zib.de/groetschel
2

About the assignment problem

The assignment problem is a mathematical
problem. Mathematicians have spent an awful lot
of time to create real-life interpretations that
look like applications to prove that it is useful.
And hence, the Hungarian Method is of no
practical value.
The truth, in fact is the other way around.
Practitioners have tuned their applied problems
in order to be able to employ the Hungarian
Method.
Martin
Grtschel
3

Contents
1. What is vehicle circulation/scheduling?
2. Single depot vehicle scheduling
3. Multiple depot vehicle scheduling
4. Extensions

Martin
Grtschel
4

Contents
1. What is vehicle circulation/scheduling?
2. Single depot vehicle scheduling
3. Multiple depot vehicle scheduling
4. Extensions

Martin
Grtschel
5

Planning Public Transportation

Phase: Planning Scheduling Dispatching

Horizon: Long Term Medium term (very) Short term

Timetable Period Day of Operation
online planning
Objective: Service Level Cost Reduction Get it done

Steps: Network Design Vehicle Scheduling Crew Assignment

Line Planning Duty Scheduling Delay Management
Timetabling Duty Rostering Failure Management

Martin
Grtschel
6

The ZIB Transportation Team,

including former members:
Public Transport: Online Transportation:
Ralf Borndrfer Norbert Ascheuer
Fridolin Klostermeier Philipp Friese
Christian Kttner Sven O. Krumke
Andreas Lbel Diana Poensgen
Sascha Lukac Jrg Rambau
Marc Pfetsch Luis Miguel Torres
Thomas Schlechte Andreas Tuchscherer
Steffen Weider Tjark Vredeveld
plus several master students
Martin
Grtschel
7

Planning in Public Transport

(Product, Project, Planned)
Operations Control

Construction Costs
Network Topology
Crew Assignment

Cost Recovery
Service Level
Relief Points

Connections
Frequencies
Disruptions

Sensitivity
Timetable
Rostering

Rotations

Velocities
Duty Mix
Fairness

Duties

Fares
Lines
IS-OPT VS-OPT2 B15
multidepartmental
Departments
multidepotwise
Depots
multiple line groups
Line Groups
multiple lines
Lines
multiple rotations
Martin
Rotations
Grtschel
APD DS-OPT VS-OPT BS-OPT AN-OPT B1 B3 B1
8

The ZIB Transportation Team

spin-off companies
Intranetz: LBW:

Fridolin Klostermeier Ralf Borndrfer

Christian Kttner Andreas Lbel
Norbert Ascheuer Steffen Weider

Martin
Grtschel
9

What is vehicle circulation/scheduling?

We are given a transportation system in a region.
It is subdivided by carrier/vehicle types (busses, trams,
subways, planes, ships).
For each carrier type, a (daily, weekly, or monthly,..)
timetable (the scheduled/timetabled trips) is given.
Task: Assign the available vehicles to the scheduled trips
of the timetable such that some objective function is
optimized and a (usually large) system of side constraints
is satisfied.

Martin
Grtschel
10

What is vehicle circulation/scheduling?

Somewhat more precise:
Each vehicle (usually) has a home base. In colloquial
language this is called its depot. Transportation
professionals have to be more precise. A depot consists
of all indistinguishable vehicles that have their home
base in the same physical location.
In most cases, a vehicle leaves its depot in the morning
and returns to its depot in the evening of the planning
period. Thus, every vehicle circulates along a tour of
the region.
The vehicle circulation problem is hence the task to find,
for each available vehicle and for the given planning
horizon, a tour such that all scheduled/timetabled trips
are covered by exactly one tour and some objective is
optimized and certain side constraints respected.
Martin
Grtschel
11

What is vehicle circulation/scheduling?

The objective function
Minimize the number of vehicles that are
necessary to cover all scheduled trips.
Minimize the cost of the deadhead trips.
(Deadhead trips are moves of a vehicle without passengers; a move can
be just a break where the vehicle keeps waiting in a parking lot.)

A combination of these two.

Interlining
Turns
Pull-in pull-out trips

Martin
Grtschel
12

Leuthardt Survey
(Leuthardt 1998, Kostenstrukturen von Stadt-, berland- und Reisebussen, DER NAHVERKEHR 6/98, pp. 19-23.)

annual cost: 150 250 thousand US dollars per bus

bus costs (DM) urban % regional %

crew 349,600 73.5 195,000 67.5
depreciation 35,400 7.4 30,000 10.4
calc. interest 15,300 3.2 12,900 4.5
materials 14,000 2.9 10,000 3.5
fuel 22,200 4.7 18,000 6.2
repairs 5,000 1.0 5,000 1.7
other 34,000 7.1 18,000 7.2
Martin total 475,500 100.0 288,900 100.0
Grtschel
13

Vehicle Scheduling in Berlin

The transportation research group at ZIB has
produced software with which the
busses
street cars, and
subways
in Berlin have been scheduled.

A film shows some of the problems of bus

scheduling:
Martin
Grtschel
14

Vihicle Circulation Film

Martin
Grtschel
15

Some Users
VS-OPT (vehicles) DS-OPT (drivers)
ATC/Terni (I) ATC/Terni (I)
Athen (U) (GR)
Berlin (D) Berlin (D)
Bonn (D) Bonn (D)
Connex (D) Connex (D)
DB Regio (D) DB Regio (D)
Geilenkirchen (D)
Ennepetal (D)
Genua (I) Genua (I)
Mailand (U) (I) Mailand (U) (I)
Mnchen (S) (D) Mnchen (S) (D)
Norgesbus (N) Norgesbus (N)
Rhein-Neckar (S) (D) Rhein-Neckar (S) (D)
Wiesbaden (D) Wiesbaden (D)
Martin
Grtschel
16

Contents
1. What is vehicle circulation/scheduling?
2. Single depot vehicle scheduling
3. Multiple depot vehicle scheduling
4. Extensions

Martin
Grtschel
17

Single Depot Vehicle Scheduling

(Assignment Model)

depot (in the morning)

D

1 2 with starting time and location

timetabled
trip
1 2 with ending time and location

4 timetabled trips
3 4
A single depot:
one location
3 4 one bus type

D depot (in the evening)

Martin
Grtschel
18

Single Depot Vehicle Scheduling

(Assignment Model)

4 timetabled trips plus

D 12 deadhead trips
1 blue and 1 red
2 depot nodes for
1 2 bus circulation
each available bus

D D 1 2 3 4
1 2

3 4 D D 1 2 3 4

The assignment model of the

3 4 single depot
vehicle circulation problem
D
Martin
Grtschel
19

But
In the seventies the available computers were
not able to solve large size assignment problems
due to time and space problems.
The Hungarian method was the algorithm of
choice. There was nothing better.

Martin
Grtschel
20

Problem Specific Size Reduction:

HOT = Hamburger OptimierungsTechnik
HOT only looked at
peak times (about 7 a.m)
and made heuristic
(manual = interactive) choices to
reduce the problem size.

~ 1975 beginning of code development

~ 2003 last installations replaced

Martin
Grtschel
21

All other companies did basically the same;

but it is hard to find out what they really did.
Martin
Grtschel
22

Surprise
Due to expertise and practical experience, the
HOT specialists were able to come up with very
good (and often almost optimal) solutions when
number of busses was the major objective.

Martin
Grtschel
23

Single Depot Vehicle Scheduling

1 2 3 Buses
4 3 6 7
3 3 7 8 10
Solution
Cost = 20

1 2 3 Trips

The Assignment Problem

Input: 3 Buses, 3 trips, costs
Output: cost minimal assignment
Martin
Grtschel
24

Single Depot Vehicle Scheduling

1 2 3 Buses
4 3 6 7
3 3 7 8 10
Solution
Cost = 17

1 2 3 Trips

The Greedy-Heuristik
heuretikos (gr.): inventive
heuriskein (gr.): to find

Martin
Grtschel
25

Single Depot Vehicle Scheduling

1 2 3 Busses
4 3 6 7
3 3 7 8 10
Solution
Cost = 16

1 2 3 Trips

The Greedy-Heuristik
heuretikos (gr.): inventive
heuriskein (gr.): to find

Martin
Grtschel
26

Single Depot Vehicle Scheduling

5 4 0 Buses
4 3 6 7
3 3 7 8 10
Optimum
Cost = 15

7 8 9 Trips

The "Primal Problem" The "Dual Problem"

Minimum Cost Maximum Sales Revenues
Assignment "Shadow Prices"

Martin
Grtschel
27

Mathematical Models
(Assignment Problem)

min 3x11 3x12 4 x13

3x21 7 x22 6 x23
7 x31 8 x32 10 x33
1 2 3 s.t. x11 x12 x13 1
4 3 6 7 x21 x22 x23 1
3 3 7 8 10
x31 x32 x33 1
x11 x21 x31 1
x12 x22 x32 1
x13 x23 x33 1
1 2 3 x11 , , x33 0
x11 , , x33 {0,1}
Graph Theoretic Model Integer Programming
Model
Martin
Grtschel
Linear Programming
28

Single Depot Vehicle Scheduling

0 0 0
4 3 6 7
3
3 4 3 7 6 7 8 10
3
3 7 8 10

0 0 0

The Successive Shortest Path Algorithm

Martin
Grtschel
29

Single Depot Vehicle Scheduling

0
0 00 0
0 0 Buses
0 0 0
4 3 6 7 Bound
3
3
3
4 3 7 6 7 8 10 cost = 15
3 7 8 10
Partial sol.
cost = 0
Trips
0 0 0
The Successive Shortest Path-Algorithm

Martin
Grtschel
30

Single Depot Vehicle Scheduling

0
0 00 0
0 0 Buses
0 +0 0 +0 0 +0
4 3 6 7 Bound
3
3
0
0 0 7 2 4 8 10 cost = 10
0 4 5 6
Partial sol.
cost = 3
Trips
0 +3 0 +3 0 +4
The Successive Shortest Path Algorithm

Martin
Grtschel
31

Single Depot Vehicle Scheduling

0
0 00 0
0 0 Buses
0 0 0
4 3 6 7 Bound
3
-3
0
0 0 7 2 4 8 10 cost = 10
0 4 5 6
Partial sol.
cost = 3
Trips
3 3 4
The Successive Shortest Path Algorithm

Martin
Grtschel
32

Single Depot Vehicle Scheduling

0
0 00 0
0 0 Buses
0 +0 0 +0 0 +0
4 3 6 7 Bound
3
-3
0
0 0 7 2 4 8 10 cost = 10
0 4 5 6
Partial sol.
cost = 6
Trips
3 +0 3 +0 4 +0
The Successive Shortest Path Algorithm

Martin
Grtschel
33

Single Depot Vehicle Scheduling

0
0 00 0
0 0 Buses
0 0 0
4 -3 6 7 Bound
-3
3
0
0 0 7 2 4 8 10 cost = 10
0 4 5 6
Partial sol.
cost = 6
Trips
3 3 4
The Successive Shortest Path Algorithm

Martin
Grtschel
34

Single Depot Vehicle Scheduling

0
0 00 0
0 0 Buses
0 +5 0 +4 0 +0
4 -3 6 7 Bound
-3
3
0
0 0 7 2 4 8 10 cost = 15
0 4 5 6
Partial sol.
cost = 15
Trips
3 +4 3 +5 4 +5
The Successive Shortest Path Algorithm

Martin
Grtschel
35

Single Depot Vehicle Scheduling

0
0 00 0
0 0 Buses
5 4 0
-4 -3 6 7 Bound
3
3 0
0
0 7 1 0 -8 10 cost = 15
1 3 0 1
Partial sol.
cost = 15
Trips
7 8 9
The Successive Shortest Path Algorithm

Martin
Grtschel
36

Single Depot Vehicle Scheduling

5 4 0 Buses
4 3 6 7 Bound
3
3 0 0 7 1 0 8 10
1
0
3 0 1 cost = 15
Solution
cost = 15
7 8 9 Trips

The Successive Shortest Path Algorithm

Path Search
Solution + Proof
Martin
Grtschel
Efficient
37

SPEC

Andreas Lbel

Martin
Grtschel
38

Contents
1. What is vehicle circulation/scheduling?
2. Single depot vehicle scheduling
3. Multiple depot vehicle scheduling
4. Extensions

Martin
Grtschel
39

Martin
Grtschel
40

Martin
Grtschel
41

Vehicle Scheduling
Input
Timetabled and deadhead trips vehicle
Vehicle types and depot capacities circulations
rotations
Vehicle costs (fixed and variable) blocks
schedules
Output
Vehicle rotations
Problem
Compute rotations to cover all timetabled trips
Goals
Minimize number of vehicles
Minimize operation costs
Martin
Minimize line hopping etc.
Grtschel
42

Graph Theoretic Model

deadhead trips
pull-out trips pull-in trips
timetabled trips

Martin
Grtschel
43

Example: Regensburg

Map deleted

Martin
Grtschel
44

Vehicle Scheduling

Depot
Fleet
No line
Peaks:
minimum:
capacities:
Turning:
changes:
pull-in/pull-out
pull-in
turns
soft
interlining
upper
tripstrips
limits
trips
Definition + cost of deadhead trips
Precise control at point, time, or trip
Changes of vehicles, lines, modes, turning, etc.
Automatic generation of pull-in/pull-out trips
Maintencance of all possible deadhead trips
Martin
Grtschel
Depot capacities (soft)
45

Timelines
d

Needs Fow model

assignment model will be too large d
Martin
Grtschel
46

Integer Programming Model

(Multicommodity Flow Problem)

min ij xij
d
c d d

ij

ij jk
x
i
d
x d

k
0 j, d Vehicle flow

ij jk
d
x d

i
x d

d k
0 j Aggregated flow

ij
d
x d

i
1 j Timetabled trips

0j
x
j
d
d d Depot capacities

xijd {0,1} ij , d Deadhead trips

Martin
Grtschel
47

Theoretical Results
Observation: The LP relaxation of the
Multicommodity Flow Problem does in general
not produce integeral solutions.
Theorem: The Multicommodity Flow Problem is
NP-hard.
Theorem (Tardos et. al.): There are pseudo-
polynomial time approximation algorithms to
solve the LP-relaxation of Multicommodity Flow
Problems which are faster than general LP
methods.
Martin
Grtschel
48

Lagrangean Relaxation
min cT x max min cT x (b Ax)

Ax b Bx d
Bx d x 0
x 0
max f ( ) max min cT xi (b Axi )
i

f3 x2
x1
f2 f4
P(A,b)P(B,d) f P(A,b)
f1 x4 x3
Martin
Grtschel
49

Bundle Method
(Kiwiel [1990], Helmberg [2000])

Max f ( ) : min c T x T (b Ax )
xX
X polyhedral (piecewise linear)

f ( ) c T x T (b Ax )
f f1
fk ( ) : min f ( )
Jk

f uk 2

k 1 argmax fk ( ) k
2 3 1 2
Martin
Grtschel
50

Primal Approximation
1
k 1 k
u J
(b Ax )
k z
fk 1
xk 1 x
Jk fk
fk ( ) c T xk (b Axk )
f
Theorem
k 1
b Axk 0 (k )

( xk )kN converges to a point x x : Ax b, x X

Martin
Grtschel
51

Quadratic Subproblem

(1) max fk ( ) uk k
2

uk
2
(2) max v k
2
s.t. v f ( ), for all Jk

2
1
(3) max f ( )
2uk
(b Ax )
Jk Jk
s.t. 1
Jk
0 1, for all Jk

Martin
Grtschel
52

Bundle Method
(IVU41 838,500 x 3,570, 10.5 NNEs per column)

450

400

350

300

250

20
0

15
0 bundle
10
volume
0 barrier
50 cascent

Martin 20 40 60 80 100 120 140 sec

Grtschel
53

Lagrangean Relaxation I

min ij xij
d
c d d

ij

ij jk
x
i
d
x d

k
0 j, d Vehicle flow

ij jk
d
x d

i
x d

d k
0 j Aggregated flow

ij
d
x d

i
1 j Timetabled trips

0j
x
j
d
d d Depot capacities

xijd {0,1} ij , d Deadhead trips

Martin
Grtschel
54

Lagrangean Relaxation I

c x ij
x
d d d d
max min ij ij 0 x jk

d ij i k
Vehicle flow
ij jk
d
x d

i
x d

d k
0 j Aggregated flow

ij
d
x d

i
1 j Timetabled trips

0j
x
j
d
d d Depot capacities

xijd {0,1} ij , d Deadhead trips

Subproblem: Min-Cost-Flow (single-depot)

Martin
Grtschel
55

Lagrangean Relaxation II

min ij xij
d
c d d

ij

ij jk
x
i
d
x d

k
0 j, d Vehicle flow

ij jk
d
x d

i
x d

d k
0 j Aggregated flow

ij
d
x d

i
1 j Timetabled trips

0j
x
j
d
d d Depot capacities

xijd {0,1} ij , d Deadhead trips

Martin
Grtschel
56

Lagrangean Relaxation II
d
c x 1 xij
d d
max min ij ij

d ij d i
ij jk
x
i
d
x d

k
0 j, d Vehicle flow

ij jk
d
x d

i
x d

d k
0 j Aggregated flow

Timetabled trips
0j
x
j
d
d d Depot capacities

xijd {0,1} ij , d Deadhead trips

Subproblem: Several independent Min-Cost-

Flows (single-depot)
Martin
Grtschel
57

Heuristics
Cluster First Schedule Second
"Nearest-depot" heuristic
Lagrange Relaxation II + tie breaker
Schedule First Cluster Second
Lagrange relaxation I
Schedule Cluster Reschedule
Schedule: Lagrange relaxation I
Cluster: Look at paths
Solve a final min-cost flow
Martin
Grtschel
Plus tabu search
58

Lagrangean Relaxation Algorithm

Martin
Grtschel
59

Computational Results

BVG HHA VHH

depots 10 14 10
vehicle types 44 40 19
timetabled trips 25,000 16,000 5,500
deadheads 70,000,000 15,100,000 10,000,000
cpu mins 200 50 28

Martin
Grtschel
60

Vehicle Utilization

Martin
Grtschel
61

"Camel Curve"

68

Martin
Grtschel
62

Interlining

too short to turn

too long to wait

best choice

Martin
Grtschel
63

Umlaufoptimierung mit MICROBUS 2

Umlaufoptimierung
Erzielte Einsparungen durch die
Umlaufoptimierung:
Im Busbereich wurden bei einer Gesamtzahl von knapp ber
200 Fahrzeugen 5 Busse eingespart.

Im Bahnbereich wurden aufgrund der fehlenden Leerfahrt-

und berholmglichkeiten keine Fahrzeuge eingespart.

Slide of SWB
Martin
Heiko Klotzbcher
Grtschel 26.02.2002 4
64

Vehicle Scheduling at ZIB

Martin
Grtschel
65

Vehicle Scheduling at ZIB

Ralf Borndrfer
Andreas Lbel

Ramifications:
Corinna Bnisch
Ines Spenke
Steffen Weider

Martin
Grtschel
66

BVG (Berlin)

Martin
Grtschel
67

Contents
1. What is vehicle circulation/scheduling?
2. Single depot vehicle scheduling
3. Multiple depot vehicle scheduling
4. Extensions

Martin
Grtschel
68

Discussion/Extensions
Properties
Exploiting all degrees of freedom
Vehicle mix
Extensions
Trip shifting current work
Multiperiod scheduling
Periodic schedules
Assimilation
Balanced depot exchange
Maintenance constraints
Integration
Vehicle and duty scheduling current work
Timetabling
Martin Line planning
Grtschel
69

Trip Shifting

Martin
Grtschel
 Vehicle Circulation and
the Hungarian Method

The END
Martin Grtschel
joint work with
Ralf Borndrfer Andreas Lbel Steffen Weider

A celebration day of the 50th anniversary

of the Hungarian Method
ThankBudapest,
you for your
October 31, attention
2005

Martin Grtschel Institute of Mathematics, Technische Universitt Berlin (TUB)

DFG-Research Center Mathematics for key technologies (MATHEON)
Konrad-Zuse-Zentrum fr Informationstechnik Berlin (ZIB)
[email protected] http://www.zib.de/groetschel