applied

sciences
Article
The MINLP Approach to Topology, Shape and Discrete Sizing
Optimization of Trusses
Simon Šilih 1, * , Zdravko Kravanja 2 and Stojan Kravanja 3

1 ANDRITZ AG, Stattegger Strasse 18, 8045 Graz, Austria

2 Faculty of Chemistry and Chemical Engineering, University of Maribor, Smetanova 17,
2000 Maribor, Slovenia; [email protected]
3 Faculty of Civil Engineering, Transportation Engineering and Architecture, University of Maribor,
Smetanova 17, 2000 Maribor, Slovenia; [email protected]
* Correspondence: [email protected]

Abstract: The paper presents the Mixed-Integer Non-linear Programming (MINLP) approach to the
synthesis of trusses. The solution of continuous/discrete non-convex and non-linear optimization
problems is discussed with respect to the simultaneous topology, shape and discrete sizing opti-
mization of trusses. A truss MINLP superstructure of different topology and design alternatives
has been generated, and a special MINLP model formulation for trusses has been developed. In the
optimization model, a mass objective function of the structure has been defined and subjected to
design, load and dimensioning constraints. The MINLP problems are solved using the Modified
Outer-Approximation/Equality-Relaxation (OA/ER) algorithm. Multi-level MINLP strategies are in-
troduced to accelerate the convergence of the algorithm. The Modified Two-Phase and the Sequential
Two-Phase MINLP strategies are proposed in order to solve highly combinatorial topology, shape
and discrete sizing optimization problems. The importance of local buckling constraints on topology
optimization is also discussed. Some simple numerical examples are shown at the end of the paper to
!"#!$%&'(! demonstrate the suitability and efficiency of the proposed method.
!"#$%&'
Citation: Šilih, S.; Kravanja, Z.; Keywords: structural synthesis; topology optimization; discrete sizing optimization; Mixed-Integer
Kravanja, S. The MINLP Approach to Non-linear Programming; MINLP; Modified OA/ER algorithm; multi-level MINLP strategies; steel
Topology, Shape and Discrete Sizing structures; trusses
Optimization of Trusses. Appl. Sci.
2022, 12, 1459. https://doi.org/
10.3390/app12031459

Academic Editor: Nikos D. Lagaros 1. Introduction

Received: 20 December 2021

Optimization of different processes and systems has become an important factor
Accepted: 26 January 2022
in every aspect of human life, especially in the fields of science, industry and business.
Published: 29 January 2022
Numerous attempts have been made in the past in order to improve the efficiency and
economic viability of structures in structural engineering. Although the beginnings of
Publisher’s Note: MDPI stays neutral
modern structural optimization go back to the end of the 19th century [1,2], the more
with regard to jurisdictional claims in
intensive growth of the topic started in the early 1960s [3], closely linked with the progress
published maps and institutional affil-
of computer capabilities. The first attempts in the field of structural optimization were
iations.
applied to solve the continuous sizing and shape optimization of structures, where all
variables were treated as continuous ones. However, at the end of the previous millennium,
many efficient optimization techniques were developed for solving the discrete/continuous
Copyright: © 2022 by the authors.
topology, shape and discrete dimension optimization. These techniques were developed
Licensee MDPI, Basel, Switzerland. particularly for the solution of real engineering structures with discrete/standard dimen-
This article is an open access article sions. Although the explicit inclusion of discrete variables in the models considerably
distributed under the terms and increases the extensiveness and the combinatoric complexity of the optimization problems,
conditions of the Creative Commons such a discrete/continuous type of the optimization usually proves to be very efficient with
Attribution (CC BY) license (https:// respect to both the obtained results and the required computational time.
creativecommons.org/licenses/by/ The present paper discusses the simultaneous topology, shape and discrete/standard
4.0/). dimension optimization of trusses using the Mixed-Integer Non-linear Programming

Appl. Sci. 2022, 12, 1459. https://doi.org/10.3390/app12031459 https://www.mdpi.com/journal/applsci

Appl. Sci. 2022, 12, 1459 2 of 36

(MINLP) approach. The general MINLP model formulation for truss structures is presented.
This discrete/continuous non-convex and non-linear optimization problem is proposed to
be solved by the Modified Outer-Approximation/Equality-Relaxation (OA/ER) algorithm
by Kravanja and Grossmann [4]. In addition, some multi-level MINLP strategies that
significantly accelerate the convergence of the optimization are presented. The suitability
and efficiency of the proposed method are shown through some numerical examples.

2. Optimization of Trusses
Trusses are probably the most frequent grid-like structures to be applied for testing
different optimization techniques since the early 1960s [3]. Compared to other types of
structures, the analysis and design processes of trusses are relatively simple and can be
easily written into a mathematical model. On the other hand, trusses are usually composed
from a large number of elements (bars) and are, therefore, unsuitable for optimization
by repeating the calculation of all different structural alternatives. In the case of real
engineering trusses, a lot of bars are oversized and thus not fully exploited. These structures
consequently offer numerous possibilities for improvement and optimization.
In general, truss optimization problems can be divided into two main spheres of
activity. The first one belongs to the topology optimization where an optimal structure with
an optional number and configuration of structural elements inside a defined superstructure
has to be defined, while the second one presents a discrete sizing optimization problem,
where the cross-sections of elements are forced to have discrete, in most cases standard,
dimensions. Both mentioned optimization spheres represent the greatest challenge to the
problem of truss optimization. In the last decades, some effective attempts of the topology
and discrete dimension optimization of trusses have been made. A brief review and a
history of different optimization techniques for solving these optimization problems are
discussed in the following sub-sections.

2.1. Topology Optimization

The first examples of truss optimization included only sizing optimization, i.e., the
calculation of optimal cross-section areas of bars. Soon after that, it was discovered that it
would be reasonable to exclude some unfavorable or needless elements from the original
truss structure. As a result, not only the optimal cross-sections would be obtained but
also the optimal number and configuration of bars, namely the optimal topology. One
of the most important contributions to the area of truss topology optimization is the so-
called ground structure approach [5], where a grid of nodes and a set of inter-connections
between nodes, representing bars, are defined. In the process of optimization, some bars
are eliminated, and the optimal topology is represented by the sub-set of existing bars.
Since trusses are analyzed using joint equilibrium equations, the set of feasible solutions
includes only statically determinate structures. Further, only stress constraints are applied,
and consequently, the problem can be formulated as a linear programming (LP) problem.
By including the deflection constraints and statically indeterminate solutions, the
discussed truss problem becomes a non-linear one. Numerous topology optimization
problems of trusses, solved by the non-linear programming (NLP) approach, were therefore
subsequently presented [6–11]. Using continuous optimization methods, i.e., the LP or
the NLP, the truss topology optimizations were performed by either allowing zero values
or by defining a small lower bound of the cross-section areas of bars. While the former
approach caused problems to be potentially insoluble due to the possible appearance of
singularities, the latter generally lead to non-exact solutions, which require reanalysis.
Moreover, allowing zero or non-zero values led to non-convexities, which usually cause
the optimization to become stuck in poor local optima.
All the mentioned difficulties led to the idea to optimize the truss topology by different
discrete optimization methods. The topology is optimized inside a discrete space by
introducing binary 0–1 variables. An extra binary variable y is associated with each truss
element, indicating whether this particular element is included in (y = 1) or excluded
Appl. Sci. 2022, 12, 1459 3 of 36

from (y = 0) the current topology. By means of discrete optimization, numerous authors

optimize the truss topology applying the genetic algorithm (GA) [12–16], the penalty
function method [10] and the simulated annealing [17–19].
The nature of the considered truss topology problem is that the optimization mod-
els are mixed discrete/continuous, with discrete variables for topology and continuous
variables for cross-section dimensions, nodal coordinates, stresses, strains, etc. As the
NLP is regarded as the most effective for solving continuous non-linear optimization prob-
lems, some attempts have been made to link the NLP with different discrete optimization
techniques. Such examples are the linkage of GA and NLP [20] and the tabu search [21].
Besides the widely used ground structure approach, a second approach to topology
optimization of trusses is the homogenization method [22–25], where the optimal material
distribution in a continuous design domain is determined and interpreted as a discrete truss.

2.2. Discrete Sizing Optimization

The first ideas for the sizing optimization of trusses by considering available discrete
dimensions were presented in the late 1960s [26]. However, as the main focus in the
1960s and 1970s was on the development of efficient algorithms for solving large-scale
NLP problems, the intense progress in discrete optimization methods did not start until
the 1980s. The first attempts of the discrete sizing optimization of trusses were based on
linear programming, e.g., Templeman’s algorithm [27,28], followed by the solving of NLP
problems with sequential linear programming (sequential linearization algorithm) [29–31].
In addition, practically all the developed discrete optimization techniques were ap-
plied to the truss problems, e.g., the genetic algorithm [32–34], simulated annealing [35,36],
the penalty function method [37,38] and neuron networks [39]. In some cases, multi-level
approaches were presented where the continuous optimization was performed at the first
level, and the obtained continuous solution was used as a starting point for the subsequent
discrete optimization, e.g., [30–32,40,41].
In most of the mentioned references, trusses were optimized by considering stress and
deflection constraints and either without or by considering the simplified Euler buckling
constraints. Cross-section areas were regarded as independent sizing variables, and a set
of discrete values of the section areas (in some cases, based on the available sections) was
defined. Taking this into account, researchers thus focused mainly on testing different
optimization techniques, while the practical applicability of the obtained results was often
kept in the background. One of the exceptions was presented in [42], where the trusses
were optimized by considering the design constraints in accordance with Eurocode 3 [43]
and by considering standard steel tubes as the cross-sections of elements.

2.3. Simultaneous Topology and Discrete Dimension Optimization

The first attempts to simultaneously optimize both the truss topology and the discrete
dimensions of members have been presented at the beginning of this century. The problem
was either simplified/linearized and was solved using the Mixed-Integer Linear Program-
ming (MILP) and the Mixed-Logic Linear Programming (MLLP) [44] or in the non-linear
form using the relative quotient algorithm [45]. In both cases [44,45], the topology was
optimized by using the previously discussed method of allowing a small lower bound on
the cross-section areas. In the next attempt, the truss problem was solved by employing
a mixed variable formulation (MILP), where truss member buckling constraints became
linear [46].
The simultaneous topology and discrete cross-section optimization of trusses was
also solved by using the NLP with heuristics [47], the relative difference quotient algo-
rithm [48], the genetic algorithm [49,50], the ant colony approach [51,52], the improved
genetic algorithm with two-level approximation [53], the mixed-integer second-order cone
programming approach [54], the integrated particle swarm optimizer [55], the modified
meta-heuristics with random mutation [56], the advanced Jaya algorithm [57], and by the
mixed integer linear optimization (MILO) [58].
Appl. Sci. 2022, 12, 1459 4 of 36

2.4. MINLP Approach to Truss Synthesis

Considering the optimization of trusses, the Mixed-Integer Non-linear Program-
ming (MINLP) approach was, in the past, applied to discrete dimension optimization
only [29–31] and was seldom used for the topology optimization. The simultaneous topol-
ogy and discrete sizing optimization represents an extensive non-convex and non-linear
discrete/continuous optimization problem that cannot be solved without suitable MINLP
algorithms and strategies. The practical applicability of the MINLP approach to structural
optimization was, for example, presented in [59,60] and employed for the simultaneous
topology and standard dimension optimization of hydraulic steel gates [61,62], single-story
steel building structures [63] and composite floor structures [64].
The present paper introduces the structural synthesis of trusses performed by the
MINLP optimization approach. Truss synthesis is performed through three steps; see
also [59]. The first is the generation of a truss superstructure for different structure/topology
and other design alternatives that are all candidates for a feasible and optimal solution.
The second is the development of a special MINLP model formulation for the defined truss
superstructure in an equation-oriented environment. The last step is the solution of the
defined MINLP model, performed by a suitable MINLP algorithm and strategies, which, in
the simultaneous MINLP optimization approach, yields an optimal solution.
As a result of the simultaneous optimization procedure, the MINLP truss synthe-
sis yields:
• the optimal structure mass or costs as the defined objective;
• the optimal truss topology with the optimal number and configuration of bars;
• the optimal joint coordinates, which define the shape of the truss; and
• the optimal discrete/standard cross-section dimensions of bars.
While the structure mass and joint coordinates represent continuous parameters,
which are optimized inside the continuous space, the optimization of topology and dis-
crete/standard dimensions require discrete decisions. The discussed truss synthesis thus
corresponds to a mixed discrete/continuous non-linear optimization problem, which can
be solved by the MINLP.

3. Development of the MINLP Model Formulation for Truss Superstructures

Truss synthesis requires the generation of an MINLP truss superstructure, which
includes all possible topology/structure alternatives to compete for a feasible and optimal
solution. The truss superstructure consists of the provided set of nodes and their inter-
connections, which represent the elements of the truss, i.e., bars. Special logical relations
between the elements have been defined in order to provide the kinematical stability of
the structure and prevent the undesirable overlapping of elements involved in different
topology/structure alternatives. In general, all nodes and elements are alternative (op-
tional), i.e., they can be selected or rejected from the structure. Some nodes, however, e.g.,
supported nodes and those where nodal loads are applied, are fixed and included in all
alternative truss designs. The fixed nodes together with some fixed elements reduce the
combinatorial expanse of the optimization problem. Besides topological alternatives, the
discrete/standard dimension alternatives are also defined in the superstructure. Although
a high number of discrete/standard alternatives, in general, leads to better results, it also
leads to extensive models and very expensive and difficult-to-solve problems.

3.1. The General MINLP Problem Formulation

The MINLP truss superstructure is modeled as an MINLP problem in which con-
tinuous and discrete variables are handled simultaneously. Continuous variables are
defined for the continuous optimization of parameters, while discrete variables are used for
discrete/standard dimensions and to express discrete decisions, i.e., the existence or non-
Appl. Sci. 2022, 12, 1459 5 of 36

existence of alternative structural elements. The general non-linear continuous/discrete

optimization problem (MINLP-G) can be formulated as follows:

min z = cT y + f (x)

s.t.

h(x) = 0

g(x)  0 (MINLP-G)

By + Cx  b

x 2 X = {x 2 Rn : xLO  x  xUP }

y 2 Y = {0,1}m

where x is a vector of continuous variables specified in the compact set X and y is a

vector of binary 0–1 variables. Functions f (x), h(x) and g(x) are continuous and differ-
entiable non-linear functions involved in the objective function, equality and inequality
constraints, respectively. Finally, By + Cx  b represents a sub-set of mixed linear equal-
ity/inequality constraints.
It should be noted that in the context of truss synthesis, continuous variables define
continuous structure parameters such as nodal coordinates, internal forces, deflections, etc.,
while binary variables represent the potential existence of structural elements and discrete
dimensions inside the truss superstructure. Equality and inequality constraints represent a
rigorous system of functions taken from structural analysis. Logical constraints that must
be fulfilled for discrete decisions and structure configurations are given by By + Cx  b. The
objective function to be minimized represents the purpose of the optimization (e.g., mass
or cost minimization) and is generally composed from a linear part cT y, which depends
on the number of selected structural elements (those with non-zero binary variables), and
from the non-linear dimension dependent part f (x).

3.2. The MINLP Model Formulation for Truss Superstructures

The above-mentioned general formulation MINLP-G was adopted for truss synthesis.
As a result, the MINLP truss superstructure formulation MINLP-TS was developed. As
a basis for the MINLP-TS formulation, the already developed continuous NLP truss op-
timization model was used. The NLP model proved successful for the continuous shape
and sizing optimization of steel trusses [65], composite trusses [66] and timber trusses by
considering joint flexibility [67].
The proposed MINLP-TS model formulation for planar truss superstructures consists
of the objective function, structural analysis and logical constraints with continuous and
binary variables:
Objective function:

n n
min MASS = r · Â Â Asf sf
i,j · si,j (1)
i =1 j=1
eli,j = 1

subjected to:
Structural analysis constraints:
Boundary constraints:

uk,i = 0, k 2 K, 8 i 2 I sx ,i = 1 (2)

vk,i = 0, k 2 K, 8 i 2 I sy ,i = 1 (3)
Appl. Sci. 2022, 12, 1459 6 of 36

Fx,k,i = px,k,i , k 2 K, 8 i 2 I sx ,i = 0 (4)

Fy,k,i = py,k,i , k 2 K, 8 i 2 I sy ,i = 0 (5)

Finite element equations:

n sf · E
Ai,j h ⇣ ⌘ ⇣ ⌘i
sf 2 · u
· li,j sf · msf · v
uk,j + li,j vk,j
 ssf k,i i,j k,i +
i,j
j=1
eli,j = 1
n Asf h ⇣ ⌘ ⇣ ⌘i k 2 K, i 2 I (6)
j,i · E
 ssf
· l sf
j,i
2· u
k,i uk,j + l sf
j,i · m sf · v
j,i k,i vk,j = Fx,k,i
j,i
j=1
el j,i = 1

n sf · E
Ai,j h ⇣ ⌘ ⇣ ⌘i
sf · msf · u
· li,j uk,j + msf 2· v vk,j
 ssf i,j k,i i,j k,i +
i,j
j=1
eli,j = 1
n Asf h ⇣ ⌘ ⇣ ⌘i k 2 K, i 2 I (7)
j,i · E
 ssf
· l sf
j,i · m sf · u
j,i k,i uk,j + msf
j,i
2· v
k,i vk,j = Fy,k,i
j,i
j=1
el j,i = 1
sf · E h
Ai,j ⇣ ⌘ ⇣ ⌘i
sf
· l i,j · uk,i uk,j + msf
i,j · vk,i vk,j = Fk,i,j , k 2 K, 8 (i, j)|el i,j = 1 (8)
ssf
i,j

Substituted functions:

Asf
i,j = f di,j , 8 (i, j ) el i,j = 1 (9)
q
sf cx cx 2 cy cy 2
si,j = j i + j i , 8 (i, j) el i,j = 1 (10)

sf c c sf
li,j = xj xi /si,j , 8 (i, j) el i,j = 1 (11)

sf c c sf
mi,j = yj yi /si,j , 8 (i, j) el i,j = 1 (12)

Design constraints

Fk,i,j  Asf
i,j ·
t,max
, k 2 K 8 (i, j) el i,j = 1 (13)

Fk,i,j Asf
i,j ·
c,max
, k 2 K 8 (i, j) el i,j = 1 (14)

Fk,i,j Asf cr
i,j · si,j , k 2 K 8 (i, j ) el i,j = 1 (15)

Deflection constraints
uk,i  uimax , k 2 K i 2 I (16)
vk,i  vimax , k 2 K i 2 I (17)
Logical constraints:
Inter-connection logical constraints:

Bytop  b (18)

Bound logical constraints:

top
di,j dUP
i,j · yi,j  0, 8 (i, j ) el i,j = 1 (19)
Appl. Sci. 2022, 12, 1459 7 of 36

top
di,j dex,LO
i,j · yi,j 0, 8 (i, j) el i,j = 1 (20)

Logical relations for common variables:

⇣ ⌘
top
di,j dcom
g dUP
i,j · 1 yi,j , 8 ( g, i, j) gel g,i,j = 1 (21)
⇣ ⌘
UP top
di,j  dcom
g + d i,j · 1 yi,j , 8 ( g, i, j) gel g,i,j = 1 (22)

Logical relations for discrete dimensions:

di,j = Â qi,j,l · ysti,j,l , 8 (i, j)|el i,j = 1 (23)

l2L

top
 ysti,j,l = yi,j , 8 (i, j)|el i,j = 1 (24)
l2L

Variables:
(dLO UP
i,j = 0)  di,j  di,j , 8 (i, j ) el i,j = 1

c LO
xi  c xi  c xUP
i , i 2 I
c LO
yi  c yi  c yUP
i , i 2 I
top
yi,j 2 {0, 1}, 8 (i, j) el i,j = 1

yst
i,j,l 2 {0, 1}, l 2 L, 8 (i, j ) el i,j = 1

A detailed description of the MINLP-TS model formulation is given in the following

sub-sections.

3.2.1. Definition of Structural Elements

The basis of the truss superstructure is composed by n defined nodes, n2N. Each
alternative truss element is defined as an inter-connection between two nodes: for each
pair of nodes i and j, where i2I = {1,2,3, . . . n} and j2J = {1,2,3, . . . n} and i < j, an element
coefficient eli,j is defined. The coefficient eli,j is valued 1, if an alternative element (bar) exists
connecting nodes i and j (in the following text denoted as i!j); otherwise, eli,j is equal to
zero. An n ⇥ n truss element matrix EL is then formed from coefficients eli,j . The definition
of a truss element and its position in the global coordinate system XY is shown in Figure 1.

3.2.2. Variables
The MINLP-TS model formulation includes continuous variables as well as discrete
binary variables. In general, continuous variables are partitioned into independent (design)
variables and into dependent (non-design) variables. In the specific case of truss synthesis,
the design variables are sizing variables (i.e., dimensions of the cross-sections of bars, di,j )
and shape variables (i.e., nodal coordinates c xi and c yi ). These variables define the structural
design. Non-design performance variables are directly dependent on the design variables
and represent the cross-section properties, internal forces, deformations, resistances of
bars, etc.
Discrete variables are represented by a vector of binary variables y = {ytop , yst }, where
sub-vectors ytop and yst stand for the topological and discrete/standard dimension binary
top
variables, respectively. In the first sub-vector, topological binary variables yi,j are defined
for all alternative structural elements involved in the superstructure. Topological binary
top
variables are subsequently partitioned into those defining nodes (yi,j , I = j, further on in
top top
the text yi,i ) and into those defining bars (yi,j , 8 (i,j) | eli,j = 1). Each topological binary
top top
variable represents the existence or selection (yi,j = 1) or non-existence or rejection (yi,j = 0)
of its associated structural element.
Appl. Sci. 2022, 12, 1459 8 of 36

Figure 1. Definition of a truss element.

The second sub-vector of binary variables contains binary variables attributed to

the discrete/standard dimensions. Let L be a set of discrete/standard alternatives of a
particular cross-section dimension. A binary variable ysti,j,l is then defined for each discrete
cross-section alternative l, l2L, for each bar element i!j.

3.2.3. Objective Function

In the present truss synthesis, the objective function Equation (1) represents the mass
of the truss superstructure, where r denotes the unit mass of the material and Asf sf
i,j and si,j
are the cross-section area and the length of element i!j, respectively.
In its simplest form, the objective function can be linear, when the cross-section areas of
bars are considered as independent sizing variables and the joint coordinates of the truss are
fixed (i.e., the lengths of the elements are fixed parameters). In the engineering structures,
however, the bars are built from standard cross-sections. Each cross-section area generally
depends on several dimensions. This is important, especially when buckling constraints are
included, as buckling resistance of a bar is dependent not only on its cross-section area but
also on the shape and individual dimensions of its cross-section. By considering circular
hollow sections (tubes), for example, each cross-section is defined by two independent
sizing variables, i.e., the tube diameter di,j and the wall thickness ti,j . The cross-section area
is therefore evaluated from the cross-section dimensions and in the constraints defined as a
substituted function Asf i,j ; see Equation (9). In addition, when the shape of the structure is
also optimized, the joint coordinates become variables, and consequently, the lengths of
the bars also become substituted functions ssf i,j ; see Equation (10). In effect, the mass of the
structure becomes a non-linear function (Equation (1a)):
n n  q
2 2
MASS = r · ⇡ · Â Â ti,j · di,j ti,j · cx
j
cx
i + cy
j
cy
i (1a)
i =1 j=1
eli,j = 1

where di,j and ti,j denote the tube diameter and the wall thickness of element i!j, respec-
tively, and c xi , c xj , c yi and c yj are the coordinates of joints i and j.
Appl. Sci. 2022, 12, 1459 9 of 36

3.2.4. Structural Analysis Constraints

Non-linear and linear parameter constraints (Equations (2)–(17)) as well as the lower
and upper bounds on continuous variables represent a rigorous system of functions taken
from structural analysis. In the proposed MINLP formulation, structural analysis con-
straints are divided into boundary constraints, finite element equations, design constraints
and deflection constraints.
The boundary constraints, Equations (2)–(5), define the supporting and loading condi-
tions of the structure. The supporting conditions are established through coefficients sx,i
and sy,i defined for each node i. The coefficient sx,i is valued 1 if node i is supported in the
direction of axis X; otherwise, it is valued 0. Similarly, sy,i is 1 when node i is supported
in Y direction. Consequently, the corresponding deflections of the supported nodes are
defined to be zero by Equations (2) and (3). The load on a truss is given by the nodal forces,
which are defined by the loading coefficients px,k,i and py,k,i in Equations (4) and (5). Each
parameter px,k,i contains the value of the nodal force, which acts on node i in the direction of
axis X and corresponds to the loading case k, k2K. In the same way, the nodal forces in the
direction of axis Y are defined by coefficients py,k,i . In the proposed MINLP-TS formulation,
the supporting and loading coefficients are arranged into a set of corresponding vectors
sx = {sx,i }, sy = {sy,i }, px,k = {px,k,i } and py,k = {py,k,i }.
The finite element equations, Equations (6)–(8), are used to calculate the reactions,
axial forces and displacements of the structure. Equations (6) and (7) are defined for each
node i. As node i can, in general, be either the starting or the ending joint of a bar, the
contributions of elements i!j (i < j; i is the starting joint) as well as j!i (j > i; i is the ending
joint) have to be considered. Parameters li,j sf and msf are the cosines of the angles between
i,j
the local longitudinal axis of bar i!j and the global axes X and Y, respectively. When
the shape of the structure is also optimized, parameters li,j sf and msf become substituted
i,j
functions of the joint coordinates; see Equations (11) and (12).
The design constraints, Equations (13)–(15), represent the calculation of the stresses
and buckling of bars, where st,max and sc,max are the tensile and compressive strength
of the material, while cr i,j is the limit buckling stress of bar i!j. It should be noted that,
according to Equation (8), the positive sign of the axial force Fk,i,j represents a tensile force.
The deflection constraints, Equations (16) and (17), ensure that the deformations of joints
remain within the allowed limits.

3.2.5. Logical Constraints

The purpose of the logical constraints (Equations (18)–(24)) is defining different dis-
crete decisions related to reducing the set of topology alternatives, defining the inter-
connections between structural elements, establishing the bounds on continuous variables
for alternative elements and linking the continuous sizing variables with discrete/standard
values. It should be noted that all the logical constraints are expressed by the use of
binary variables.
1. Inter-connection logical constraints. Integer equality/inequality constraints, see
Equation, (18), are proposed to provide the kinematical stability of the structure,
to prevent the overlapping of elements, etc. They are provided by the defining
of simple relations between the corresponding topological binary variables. The
inter-connection constraints can be divided into bar–bar constraints and joint–bar
constraints. Some general examples of bar–bar constraints are:
top top
• bar i!j can be selected only if bar i’!j’ is also selected: yi,j yi0 ,j0  0;
top top
• if bar i!j is selected, bar i’!j’ must also be selected: yi,j yi0 ,j0 = 0;
• bars i!j and i’!j’ exclude each other, i.e., only one can be selected at the most:
top top
yi,j + yi0 ,j0  1, etc.
An example of a joint–bar constraint is:
Appl. Sci. 2022, 12, 1459 10 of 36

• bar i!j can be selected only if both of its end joints (i and j) are selected:
top top top
2 · yi,j  yi,i + y j,j .
2. Bound logical constraints. Equations (19) and (20) are proposed to define the bounds
of individual sizing variables di,j and enforce these variables to zero values if their
top
associated bar does not exist (the associated topological binary variable yi,j = 0).
Otherwise, the variables are subjected to the lower bounds dex,LO
i,j and to the upper
top
bounds dUP
i,j when the alternative bar exists (yi,j = 1). It should be noted that the
values of dex,LO
i,j are strictly larger than zero and should be distinguished from the
actual lower bounds dLO i,j , which must be zero in the case when the associated bars
disappear. In this way, the cross-sectional areas of inactive bars are zero, and such bars
do not contribute to the stiffness matrix of the structure and the objective function.
3. Logical relations for common variables. Although a truss is usually fully exploited when
each bar has a different cross-section, it is nevertheless (e.g., for construction reasons) ad-
vantageous for some groups of bars to have equal cross-sections. Equations (21) and (22)
assure that a cross-section dimension di,j of each bar i!j, which belongs to group
top
g, g2G, takes the value dcom
g when the bar is selected (yi,j = 1). Otherwise, when
top
yi,j = 0, the design variables take zero values due to Equations (19) and (20), while
Equations (21) and (22) become redundant. The linkage of bars to certain groups is
defined through the coefficients gelg,i,j , which take the value 1 if bar i!j belongs to
group g; otherwise, they are set to value 0.
4. Logical relations for discrete/standard dimensions. The linear equality constraints,
Equations (23) and (24), assure that when an element exists, the cross-section dimen-
sion di,j takes one of the defined alternative discrete/standard dimension values. For
each bar i!j, a set of l, l2L, alternative discrete dimension values qi,j,1 are defined,
and to each qi,j,1 , an extra binary variable yst i,j,l is assigned. The dimension di,j is
then defined as a scalar product between a vector of the alternative discrete dimen-
sion values qi,j,l = {qi,j,1 , qi,j,2 , . . . qi,j,l } and the associated vector of binary variables
n o
yst st st st
i,j,l = yi,j,1 , yi,j,2 , . . . , yi,j,l ; see Equation (23). Additionally, by Equation (24), only
top
one discrete value can be associated to a selected bar (yi,j = 1) and none to a rejected
top
bar (yi,j = 0).

4. Solution of the MINLP Truss Synthesis Problem

After the MINLP model formulation has been developed, the defined MINLP synthesis
problem is solved by the use of a suitable MINLP algorithm. A general MINLP class of
optimization problems (MINLP-G) can, in principle, be solved by the following algorithms
and their extensions:
• Non-linear Branch and Bound method, NBB [68,69];
• LP/NLP-based Branch and Bound algorithm [70];
• Sequential Linear Discrete Programming, SLDP [30];
• Extended Cutting Plane method, ECP [71];
• Generalized Benders Decomposition [72,73];
• Outer-Approximation algorithm, OA [74];
• Feasibility Technique [75];
• Outer-Approximation/Equality-Relaxation algorithm, OA/ER, [76].
From a variety of different techniques and their individual characteristics, the OA/ER
algorithm has proved to be very successful in solving large-scale MINLP problems in cases
when NLP sub-problems are expensive and difficult to solve. As the MINLP problems of
truss synthesis are highly non-linear, the OA/ER algorithm was selected to fulfil this task.
Appl. Sci. 2022, 12, 1459 11 of 36

4.1. The Modified OA/ER Algorithm

The OA/ER algorithm comprises the solving of an alternative sequence of Non-linear
Programming (NLP) optimization sub-problems and Mixed-Integer Linear Programming
(MILP) master problems; see Figure 2. The former corresponds to the continuous optimiza-
tion of parameters for a superstructure with fixed discrete/binary variables and yields an
upper bound to the objective to be minimized. The latter involves a global approximation
to the superstructure of alternatives in which new values of discrete/binary variables
(new topology and standard dimensions) are identified so that its lower bound does not
exceed the current best upper bound. A global linear approximation includes the linear
constraints from the original MINLP problem as well as the linearizations of the non-linear
objective function and the non-linear (in)equality constraints, accumulated at each NLP
sub-problem solution.

Figure 2. Steps of the Outer-Approximation/Equality-Relaxation algorithm.

Appl. Sci. 2022, 12, 1459 12 of 36

The OA/ER algorithm, like all other MINLP methods, does not generally guarantee
that the obtained solution is the global optimum. This is due to the presence of non-
convex functions that may cut off the global optimum. In order to reduce the undesirable
effects of non-convexities, a Modified OA/ER algorithm was developed by Kravanja and
Grossmann [4]. The Modified OA/ER algorithm allows the following modifications to
be applied to the master problem: deactivation of the linearizations, decomposition and
deactivation of the objective function linearization, use of the penalty function, use of the
upper bound on the objective function to be minimized as well as a global convexity test
and the validation of the outer approximations; see also [59].

4.2. Multi-Level MINLP Strategies

Besides many non-linearities and non-convexities, a very high number of discrete
variables, particularly those associated with discrete dimensions, may be included in the op-
timization. The simultaneous topology and standard dimension optimization can generally
be performed in a single MINLP phase. With the Single-Phase MINLP approach [77], the
optimal result is reached directly in a single optimization process together with the optimal
topology and standard dimension. All the binary variables included in the optimization
are initialized in a single full set. The disadvantage of this approach is that the initialization
scheme is weak, since the chosen initial standard dimensions may be infeasible or a bad
choice. This strategy also exhibits slow convergence when applied to large-scale MINLP
problems, and the result obtained may be a poor local solution. In order to overcome
these problems, we applied multi-level MINLP [60,77–80]. From a number of multi-level
strategies, two were applied to truss synthesis:
1. The Two-Phase (TP) MINLP approach [80], and
2. The Linked Two-Phase (LTP) MINLP strategy [60].
The TP approach performs topology, shape and standard dimension optimization
separately in two phases. In the first phase, the simultaneous topology, shape and con-
tinuous sizing optimization is performed with the dimensions of the cross-sections being
temporarily relaxed into continuous parameters. When the optimal topology is obtained,
the discrete dimensions of cross-sections are re-established, and the process continues with
the second phase, where the shape and standard dimension optimization is performed until
the optimal solution is obtained. In the second phase, the optimization is carried out at
the fixed optimal topology obtained in the first phase. The TP approach is thus performed
through a hierarchic decomposition of discrete binary variables into two sub-sets. The first
sub-set is used to describe topology alternatives and the second one to describe discrete
dimension alternatives.
The main disadvantage of the TP strategy lies in the fact that the final solution is not
necessarily the optimal one since the topology and standard dimensions are optimized
consecutively. A simultaneous consideration of the standard dimensions may cause a
change in the optimal topology [60]. In order to avoid the disadvantages of the TP approach,
the LTP strategy was developed. While the first phase of the LTP strategy is identical to
the first phase of the TP approach, the second phase differs in the fact that by using the
LTP strategy, the topology is not fixed. The second phase of the LTP strategy thus performs
simultaneous topology, shape and standard dimension optimization.
As the majority of binary variables is attributed to discrete/standard dimensions,
the design space in the previously described second phases is still very extensive, and in
some cases, even when the TP strategy is used, a very high number of MINLP iterations
is needed before the first feasible solution is attained. In order to additionally reduce the
discrete space, after the first phase is complete (topology optimization with continuous
sizing variables), a special pre-screening of the discrete dimensions is applied. With the
use of the pre-screening procedure the discrete solution is sought only in some pre-defined
neighborhood of the continuous solution obtained in the first phase.
Appl. Sci. 2022, 12, 1459 13 of 36

4.3. Pre-Screening of Discrete/Standard Dimensions

In order to reduce the number of binary variables attributed to discrete dimensions and
considering the fact that the optimal discrete dimensions are expected to be near the optimal
continuous dimensions obtained in the previous phase, the pre-screening technique was
applied after the first phase (after obtaining the optimal topology and continuous sizes).
The alternative discrete dimension values in the vector qi,j,l = {qi,j,1 , qi,j,2 , . . . qi,j,l } from
Equation (23) are positioned from the smallest discrete value in the first position to the
ps
largest value in the last position. A special parameter li,j is then defined for each bar i!j
at the m–th MINLP iteration (MINLPm , m2M), as a qi,j,l vector’s position counter l, l2L,
m 1
where the continuous dimension di,j , obtained in the previous NLPm 1 sub-problem,
is positioned next to its discrete standard value qi,j,l ; see Equation (25). The subsequent
st for standard dimensions is for the current MILPm
initialization of binary variables yi,j,l
master problem, then performed by Equation (26):

ps m 1
li,j = l , qi,j,l  di,j < qi,j,l +1 , 8 (i, j) el i,j = 1 (25)

ps
yst
i,j,l = 0 if l li,j > nps+
ps
0  yst
i,j,l  1 if nps  l li,j  nps+ (26)
ps
yst
i,j,l =0 if l li,j < nps
m 1
where nps and nps+ , close to the previously obtained continuous dimension di,j , define
the number of neighboring lower and upper discrete dimensions, which will take an active
part in the subsequent discrete sizing optimization. The binary variables lying inside the
neighborhood limited by nps and nps+ are hence active (0  yst i,j,l  1), while the rest of
them are deactivated, i.e., set to zero (yst
i,j,l = 0). In this way, only the reduced set of active
binary variables is involved in the optimization that significantly decreases the discrete
combinatorial problem.

4.4. The Proposed MINLP Strategies for Truss Synthesis

Both the TP and LTP multi-level MINLP strategies were applied to truss synthesis
problems. The TP approach yielded good solutions in a few MINLP iterations. In the use
of the LTP strategy, however, no feasible solutions were obtained in reasonable CPU times.
The search space in the second phase proved to be too extensive. Moreover, due to the
nature of truss structures, the pre-screening of standard dimensions cannot be performed
effectively because only elements selected from the first phase topology can provide useful
information for the subsequent second phase.
As mentioned before, by using the Two-Phase MINLP strategy, the optimal topology
obtained in the first phase (topology and continuous sizing optimization) is adopted as the
final optimal topology. The possible change in topology due to the inclusion of discrete
sizing variables in the second phase is thus prevented. In order to allow the topology to be
changed after the activation of discrete sizing variables and in order to gain information of
the influence of their inclusion on the subsequent optimization, some modifications were
introduced into the TP strategy. Two varieties of the TP strategy are thus proposed: the
Modified Two-Phase MINLP strategy and the Sequential Two-Phase MINLP strategy. Both
strategies are described in the following sub-sections.

4.4.1. The Modified Two-Phase (MTP) MINLP Strategy

The Modified TP strategy (see Figure 3) starts (at n = 1) with topology, shape and
sizing optimization where cross-section dimensions are relaxed into continuous parameters
(MINLP phase T). The process proceeds until the convergence of the OA/ER algorithm
is achieved, i.e., when the lower bound, obtained at the MILP master problem, exceeds
the upper bound yielded by the best NLP sub-problem. The obtained topology (with
the theoretical value of the objective ztn ) is adopted as the optimal topology of the first
Appl. Sci. 2022, 12, 1459 14 of 36

phase. After that, the pre-screening of discrete dimensions for selected bars is applied. The
calculation process proceeds with the second phase (MINLP phase D), where the shape
and discrete sizing optimization is performed at the fixed optimal topology from the first
phase. The obtained solution (with the discrete value of the objective zdn ) now represents
the structure with optimal topology, shape and discrete/standard values of cross-section
dimensions (zOPT = zdn ). Up to this stage, the process coincides with the classical Two-Phase
MINLP approach. However, by terminating the process at this point, one would not
gain any information about the possible influence of the incorporation of discrete sizing
variables on the subsequent topology optimization.

Figure 3. The Modified Two-Phase (MTP) MINLP strategy.

The process is therefore returned back to the first phase with the topology optimization
in which the discrete sizing variables are again relaxed into continuous parameters. At
this stage, the accumulated linearizations of the previous discrete sizing optimization
phase are also included in the optimization. As the »integer cut« equations prevent all
topologies obtained in previous iterations from being re-calculated, former topologies are
now excluded from the superstructure, which may cut off some good solutions. For this
reason, all the integer cuts with the exception of the one resulting from the optimal topology
are thus deactivated after the conclusion of the initial Two-Phase procedure.
In the continuation (n = n + 1), each time a topology is obtained, which would yield
a theoretical objective value ztn better than the current optimal solution (ztn < zOPT ), a suc-
cessive discrete sizing optimization at the fixed current n-th topology is performed. The
process proceeds until a defined number (nfin ) of subsequent new topologies, which do
Appl. Sci. 2022, 12, 1459 15 of 36

not fulfil the criterion ztn < zOPT , is reached. The number nfin is defined on the basis of
the estimation of the non-convexity problem, when the objective is not improved strictly
monotonically, and a new better solution is expected to be gained after some worse inter-
mediate solutions. However, a small number (i.e., from two to five) of MINLP iterations,
which do not yield an improvement in the objective, is generally sufficient to terminate the
optimization process.

4.4.2. The Sequential Two-Phase (STP) MINLP Strategy

The Sequential Two-Phase (STP) MINLP (see Figure 4) strategy similarly to the MTP
strategy consists of running a sequence of alternating topology optimization phases with
continuous sizing variables and discrete sizing optimization phases (for better topologies
only, ztn < zOPT ). The main difference between these two strategies is that the STP strategy
starts with the initial phase (n = 1), where all alternative elements are active. The initial
topology is thus the maximal topology. The process starts with the shape and continuous
sizing optimization in the initial maximal topology, followed by the pre-screening of
discrete dimensions. The subsequent discrete sizing optimization yields the first discrete
solution with the value of the objective zd1 . This solution is adopted as the currently best
solution (zOPT = zd1 ).

Figure 4. The Sequential Two-Phase MINLP strategy.

Further optimization cycles are then carried out identically according to MTP strategy,
but without the deactivations of integer cuts. The discrete sizing variables are then relaxed
into continuous parameters. The process proceeds (n = n + 1) with the topology and shape
optimization searching for a topology that would yield a solution (value of the objective
ztn ) better than the currently best discrete solution, i.e., a solution which would satisfy
the condition ztn < zOPT . Only when the condition is satisfied is the subsequent discrete
Appl. Sci. 2022, 12, 1459 16 of 36

sizing optimization at the fixed current topology performed, yielding the solution zdn . Two
successive phases, i.e., the topology optimization and the discrete sizing optimization,
together form the Two-Phase (TP) cycle. When the discrete solution zdn is obtained, the
current TP cycle is concluded, and the next cycle begins. The process is terminated if a new
topology satisfying the condition ztn < zOPT is not obtained in the defined number (nfin ) of
successive MINLP iterations.

5. Numerical Example: Synthesis of a Ten-Bar Truss Cantilever

The MINLP topology, shape and discrete dimension optimization of a truss cantilever
is presented as an illustrative example of truss synthesis. The ten-bar truss cantilever, see
Figure 5, has frequently been used to test different optimization techniques on different
optimization levels and, recently, has also been used for combined topology and discrete
sizing optimization; see, for example, references [47–50].

Figure 5. The ten-bar truss cantilever.

5.1. Input Data

The input data are taken from [49] and are as follows (see also Figure 5): L = 914.4 cm,
P = 445.4 kN, Young’s modulus E = 68.97 GPa, limit stresses t,max = c,max = 172.4 MPa and
the allowed vertical displacements of the unsupported joints vmax i = 5.08 cm. The objective
of the optimization is to minimize the weight of the structure. The term r in the objective
function thus represents the unit weight of material 27126.4 N/m3 .
In the present example, the independent sizing variables are the cross-sectional areas
of bars. A vector of l, l 2 L = {1,2,3, . . . 16}, alternative discrete values for cross-sectional
areas, is defined as follows:

qi,j,l = {6.45, 19.35, 32.26, 51.61, 64.51, 67.74, 77.42, 96.77, 109.68, 141.94, 154.84, 167.74,
180.64, 187.10, 200.00, 225.81} [cm2 ].

5.2. The Superstructure and Topological Alternatives

The truss cantilever is composed of 6 joints and 10 joint inter-connections—bars.
The basis of the superstructure is composed from identical sets i and j, i2I, j2J, where
Appl. Sci. 2022, 12, 1459 17 of 36

I ⌘ J = {1,2,3,4,5,6}. Regarding the defined bars in Figure 5, the inter-connection matrix

EL is: 2 3
0 0 1 1 0 0
6 0 0 1 1 0 0 7
6 7
⇥ ⇤ 6 0 0 0 1 1 1 7
EL = eli,j = 66 0 0
7
7
6 0 0 1 1 7
4 0 0 0 0 0 1 5
0 0 0 0 0 0
Since each nodal load represents a separate load case, there are two load cases defined,
k2K = {1,2}. According to the MINLP-TS formulation, the load vectors are defined as px,1,i
⌘ px,2,i = {0,0,0,0,0,0}, py,1,i = {0,0, P,0,0,0} and py,2,i = {0,0,0,0, P,0}. Joints 1 and 2 are sup-
ported in both X and Y directions; thus, the supporting vectors are sx,i ⌘ sy,i = {1,1,0,0,0,0}.
top top top top
A single topological binary variable yi,i is defined for each of the six joints (y1,1 ,y2,2 ,y3,3 ,
top top top top
y4,4 ,y5,5 ,y6,6 ), and a single topological binary variable yi,j is subjected to each of the ten
top
bars, i.e., to each pair of joints i and j, for which the inter-connection coefficient eli,j = 1 (y1,3 ,
top top top top top top top top top top
y1,4 ,y2,3 ,y2,4 ,y3,4 ,y3,5 ,y3,6 ,y4,5 ,y4,6 ,y5,6 ). The remaining topological binary variables yi,j are
not active and are set to zero values. Binary variables are also defined for discrete sizing
alternatives. Regarding the 16 different alternative discrete values from the defined vector
qi,j,l , an extra binary variable yst i,j,l is defined for each l-th discrete value for each i!j bar. For
the 10 defined bars, the considered problem thus involves 10 ⇥ 16 = 160 binary variables
for discrete dimensions. In total, 6 + 10 + 160 = 176 binary variables are defined.
With respect to the supporting and loading of the structure as well as the arrangement
of bars, joints 1 to 5 are fixed joints and will be included in all possible topology alternatives,
while joint 6 is an alternative joint, since it may be included in or removed from the
superstructure. In addition, bars 1!3, 3!5 and 2!4 are fixed bars, while all other
bars represent alternative structural elements. Regarding the requirement for kinematical
stability of the structure, a set of inter-connection logical constraints (18) for the present
example can be written using Equations (18a–d) as follows:
top top top
y1,4 + y2,3 + y3,4 2 0 (18a)

top top top top

y3,6 + y4,5 + y5,6 y6,6 1 0 (18b)
top top
yi,6 y6,6  0, i 2 I, 8 (i, 6) el i,6 = 1 (18c)
top top
y4,6 y6,6 = 0 (18d)
Equation (18a) concerns the panel 1–2–3–4 in which at least two bracing bars, i.e.,
one of the diagonals (1!4 or 2!3) and the vertical (3!4), or just both diagonals or all
three of them have to be selected in order to provide the kinematical stability. The possible
arrangements thus define four different topological alternatives for panel 1–2–3–4. In panel
3–4–5–6, joint 6 can be removed; therefore, according to Equation (18b), at least two bracing
top
members have to be present when joint 6 is selected (y6,6 = 1), and a single bracing member
top
is determined if joint 6 is rejected (y6,6 = 0). Equation (18c) calculates that all bars attached
to joint 6 (i.e., bars 3!6, 4!6 and 5!6) will be rejected in the case when joint 6 is rejected,
while bar 4!5 will be selected by the Equation (18b). Finally, bar 4!6 will always be
selected if joint 6 is selected; see Equation (18d).
In the panel 3–4–5–6, 4 topological alternatives therefore exist when joint 6 is selected
(equivalently to panel 1–2–3–4), plus one supplementary alternative when joint 6 is rejected,
i.e., together, five different topological alternatives. The defined superstructure of the truss
cantilever thus involves all the defined alternative combinations between the first and
the second panel, i.e., 4 ⇥ 5 = 20 topological alternatives in total. The superstructure is
schematically presented in Figure 6, where solid lines represent fixed structural elements
Appl. Sci. 2022, 12, 1459 18 of 36

(joints and bars), while dashed lines denote alternative ones. It should be noted that eight
topological binary variables are subjected to the defined eight fixed structural elements
(five joints and three elements), while the other eight binary variables are attributed to the
eight alternative structural elements. During the optimization, the values of the binary
variables, which are subjected to the fixed elements, are fixed to 1. Thus, the total number
of binary variables active in the optimization process is 176 8 = 168.
It should be noted that in the case of seven alternative bars, the total number of all
possible topological arrangements can be calculated as 27 = 128. By including heuristics
and by using simple logical inter-connection equations mentioned before, a considerable
reduction of topological alternatives is made possible and, thus, also the combinatorial
expanse of the problem.
The independent sizing variables are the cross-sectional areas of bars. Each cross-
section area is defined by a single sizing design variable di,j , and the substituted function
Equation (9) in the formulation MINLP-TS is simply Asf i,j = di,j .

Figure 6. Fixed and alternative structural elements.

5.3. The MINLP Syntheses

Using the developed MINLP-TS formulation, the following examples of synthesis of
the ten-bar truss cantilever were performed:
1. Example TC10a: Topology and discrete dimension optimization without buckling constraints.
The structure is optimized at a fixed shape (at fixed nodal coordinates) under stress
and displacement constraints, while the buckling constraints Equation (15) are ex-
cluded. The independent variables are the cross-sectional areas of bars Ai,j , with the
lower/upper bounds defined to be equal to the smallest and the largest discrete value
of vector qi,j,l (Aex,LO
i,j = 6.45 cm2 , AUP 2
i,j = 225.81 cm ).
2. Example TC10b: Topology, shape and discrete dimension optimization without buckling
constraints. Alongside the sizing variables Ai,j , the joint coordinates (c xi , c yi ), too,
are now the independent variables. Considering the supporting and loading con-
ditions, the coordinates of joints 4 and 6 are changeable, while the other joints are
fixed. The bounds on shape variables are: c x4LO = 114 cm, c x4UP = 1914 cm, c yLO 4 =
c yLO = 214 cm, c yUP = c yUP = 1214 cm, c xLO = 1028 cm, c xUP = 1928 cm.
6 4 6 6 6
Appl. Sci. 2022, 12, 1459 19 of 36

3. Example TC10c: Topology and discrete dimension optimization including buckling con-
straints. Buckling constraints Equation (15) are added to the example TC10a. The
limit buckling stress of a compressed element is considered to be equal to the Euler
cr = A ⇡E/ (4 ssf 2 ). The buckling
buckling stress for circular cross-sections; thus, si,j i,j i,j
lengths of elements are considered to be equal to the system lengths for both in-plane
and out-of-plane buckling. Taking into consideration the buckling constraints, the
cross-sectional areas of bars are expected to take higher values. The vector of discrete
values is thus expanded to:

qi,j,l = {6.45, 19.35, 32.26, 51.61, 64.51, 67.74, 77.42, 96.77, 109.68, 141.94, 154.84, 167.74,
180.64, 187.10, 200.00, 225.81, 245.14, 270.94, 290.30, 309.65, 322.55, 354.81, 387.06,
419.25} [cm2 ].

The new vector contains 24 alternative discrete values, and the total number of active
binary variables is 10 ⇥ 24 + 8 = 248.
4. Example TC10d: Topology, shape and discrete dimension optimization including buckling
constraints. The bounds on shape variables are the same as defined in example TC10b,
while the buckling constraints and the vector of discrete cross-sectional areas are the
same as in example TC10c.
All four examples of truss synthesis were carried out using both the MTP and the
STP strategies in order to find and estimate the advantages of each MINLP strategy. In
all the performed optimizations, the process was terminated after the theoretical result
ztn of topology optimization with continuous sizing variables had not improved in five
successive MINLP iterations; thus, nfin = 5.
When pre-screening of binary variables for discrete dimensions was used, two neigh-
boring (one lower and one upper) discrete values to the corresponding continuous value
obtained in the previous continuous sizing optimization were active in the successive
discrete sizing optimization. All other discrete values were temporarily deactivated, i.e.,
their associated binary variables were given zero value. The tests showed that in the case of
discrete sizing optimization at fixed topology, the first feasible solution always represented
the optimal result, while all following feasible solutions represented worse results. Thus,
each discrete sizing optimization phase was concluded when the first feasible solution
was obtained.
As an interface for mathematical modeling and data inputs/outputs GAMS (Gen-
eral Algebraic Modeling System) [81] was used. The syntheses were carried out by a
user-friendly version of the MINLP computer package MIPSYN [78,82], the successor
of PROSYN [4] and TOP [59–61]. MIPSYN is the implementation of many advanced
optimization techniques, most important of which are the Modified OA/ER algorithm
and multi-level MINLP strategies. GAMS/CONOPT [83] (generalized reduced gradient
method) was used to solve NLP sub-problems, and GAMS/CPLEX [84] (Branch and Bound)
was used to solve MILP master problems.
The convergences of the Modified OA/ER algorithm and the multi-level strategies
applied to different examples of the synthesis of trusses are shown in Tables 1–4 for
problems without buckling and in Tables A1–A4 in Appendix A for problems with buckling.
Since the initial phase of the MTP strategy corresponds to topology optimization with
continuous sizing variables, the convergences to the optimal topologies are also shown
graphically in Figures 7 and 8 for problems without buckling and in Figures A1 and A2 in
Appendix A for problems with buckling.
Appl. Sci. 2022, 12, 1459 20 of 36

Table 1. Convergence of the Modified OA/ER algorithm and the MTP strategy, example TC10a
(topology optimization, without buckling constraints).

Optimization Value of Objective Function

MINLP Iteration No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
1 1. NLP 19,164.1340 - 0.102
2 1. MILP 12,604.1900 8 0.030
2. NLP 23,425.0739 0.090
3 2. MILP 17,669.4100 0.030
T 3. NLP 19,093.9919 0.070
4 3. MILP 18,682.2600 0.040
n=1
4. NLP 19,768.0154 0.117
5 4. MILP 19,340.2000 0.040
5. NLP 19,010.8946 (zt ) 0.102
... ... ... 14
D 60 59. MILP 19,266.5400
60. NLP 19,266.5406 (zOPT ) (*) 9.036
61 60. MILP 12,627.5500 8 0.050
61. NLP 23,425.0739 0.098
62 61. MILP 15,732.7000 0.060
T
62. NLP 19,768.0154 0.078
n=2 63 62. MILP 15,826.6000 0.060
63. NLP 19,093.9919 (zt < zOPT ) 0.109
... ... ... 12
D 90 89. MILP 19,559.6700
90. NLP 19,559.6663 (zd > zOPT ) (*) 5.320
91 90. MILP 22,588.0650 8 0.070
91. NLP 23,811.3197 (z1 t > zOPT ) 0.078
92 91. MILP 27,918.6900 0.110
92. NLP 20,355.3448 (z2 t > zOPT ) 0.090
93 92. MILP 29,147.8200 0.100
n=3 T 93. NLP 19,480.2377 (z3 t > zOPT ) 0.152
94 93. MILP 32,893.4400 0.110
94. NLP 19,397.1404 (z4 t > zOPT ) 0.148
95 94. MILP 36,719.7900 0.200
95. NLP 23,757.5489 (z5 t > zOPT ) 0.160
 = 16.650
(*)
sum of CPU times for the entire D phases; MINLP phase T: Topology and continuous sizing optimization;
MINLP phase D: Discrete sizing optimization at fixed topology.

Table 2. Convergence of the Modified OA/ER algorithm and the STP strategy, example TC10a
(topology optimization, without buckling constraints).

Optimization Value of Objective Function

MINLP Iteration No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
T 1 1. NLP 19,164.1340 (zt ) - 0.102
... ... ... 20
n=1
D 39 38. MILP 19,492.8000
39. NLP 19,492.7981 (zOPT ) (*) 4.981
40 39. MILP 12,208.5400 8 0.030
40. NLP 23,425.0739 0.082
T
41 40. MILP 15,354.2100 0.040
n=2 41. NLP 19,093.9919 (zt < zOPT ) 0.078
... ... ... 12
D 68 67. MILP 19,559.6700
68. NLP 19,559.6663 (zOPT ) (*) 4.063
Appl. Sci. 2022, 12, 1459 21 of 36

Table 2. Cont.

Optimization Value of Objective Function

MINLP Iteration No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
69 68. MILP 15,589.9800 8 0.030
69. NLP 19,768.0154 0.098
T
70 69. MILP 16,178.2200 0.050
n=3 70. NLP 19,010.8946 (zt < zOPT ) 0.133
... ... ...
D 125 124. MILP 19,266.5400 14
125. NLP 19,266.5406 (zOPT ) (*) 12.236
126 125. MILP 16,999.0730 8 0.060
126. NLP 23,811.3197 (z1 t > zOPT ) 0.082
127 126. MILP 21,018.2310 0.140
127. NLP 20,355.3448 (z2 t > zOPT ) 0.102
128 127. MILP 22,321.5590 0.090
n=4 T 128. NLP 19,480.2377 (z3 t > zOPT ) 0.102
129 128. MILP 21,953.1130 0.130
129. NLP 19,397.1404 (z4 t > zOPT ) 0.188
130 129. MILP 23,778.8850 0.140
130. NLP 23,757.5489 (z5 t > zOPT ) 0.191
 = 23.148
(*)
sum of CPU times for the entire D phases; MINLP phase T: Topology and continuous sizing optimization;
MINLP phase D: Discrete sizing optimization at fixed topology.

Table 3. Convergence of the Modified OA/ER algorithm and the MTP strategy, example TC10b
(topology and shape optimization, without buckling constraints).

Optimization Value of Objective Function

MINLP Iteration No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
1 1. NLP 17,025.3822 - 0.441
2 1. MILP 13,466.9650 8 0.020
2. NLP 17,919.2246 1.020
T 3 2. MILP 16,313.8410 0.030
3. NLP 17,346.7775 2.000
n=1
4 3. MILP 17,148.3880 0.040
4. NLP 16,767.5972 (zt ) 1.020
... ... ... 16
D 11 10. MILP 15,609.8720
11. NLP 16,936.2314 (zOPT ) (*) 9.380
12 11. MILP 16,054.3400 8 0.040
12. NLP 17,919.2246 (z1 t > zOPT ) 2.469
13 12. MILP 17,259.0410 0.050
13. NLP 17,346.7775 (z2 t > zOPT ) 0.492
14 13. MILP 18,156.3480 0.060
n=2 T 14. NLP 16,978.6881 (z3 t > zOPT ) 2.137
15 14. MILP 22,324.8906 0.060
15. NLP 19,004.3514 (z4 t > zOPT ) 0.449
16 15. MILP 26,279.0820 0.080
16. NLP 18,044.2047 (z5 t > zOPT ) 0.203
 = 19.991
(*)
sum of CPU times for the entire D phases; MINLP phase T: Topology and continuous sizing optimization;
MINLP phase D: Discrete sizing optimization at fixed topology.
Appl. Sci. 2022, 12, 1459 22 of 36

Table 4. Convergence of the Modified OA/ER algorithm and the STP strategy, example TC10b
(topology and shape optimization, without buckling constraints).

Optimization Value of Objective Function

MINLP Iteration No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)

T 1 1. NLP 17,025.3822 (zt ) - 0.430

... ... ... 20
n=1
D 8 7. MILP 15,967.6240
8. NLP 17,168.1830 (zOPT ) (*) 2.550
9 8. MILP 13,466.9700 8 0.030
9. NLP 17,919.2246 1.020
10 9. MILP 16,313.8410 0.030
T
10. NLP 17,346.7775 2.000
n=2 11 10. MILP 17,148.3880 0.040
11. NLP 16,767.5972 (zt < zOPT ) 1.031
... ... ... 16
D 18 17. MILP 15,609.8720
18. NLP 16,936.2314 (zOPT ) (*) 9.334
19 18. MILP 19,935.8230 8 0.050
19. NLP 16,978.6881 (z1 t > zOPT ) 2.160
20 19. MILP 19,949.4000 0.051
20. NLP 19,004.3514 (z2 t > zOPT ) 1.309
21 20. MILP 21,562.9020 0.060
n=3 T 21. NLP 18,044.2047 (z3 t > zOPT ) 0.199
22 21. MILP 25,832.8800 0.070
22. NLP 18,831.2099 (z4 t > zOPT ) 0.563
23 22. MILP 28,257.8040 0.090
23. NLP 17,898.1026 (z5 t > zOPT ) 0.461
 = 21.478
(*)
sum of CPU times for the entire D phases; MINLP phase T: Topology and continuous sizing optimization;
MINLP phase D: Discrete sizing optimization at fixed topology.

Figure 7. Convergence to the optimal topology, example TC10a (initial phase of the MTP strategy),
example TC10a (topology optimization, without buckling constraints).
Appl. Sci. 2022, 12, 1459 23 of 36

Figure 8. Convergence to the optimal topology (initial MINLP phase of the MTP strategy), example
TC10b (topology and shape optimization, without buckling constraints).

5.4. Discussion of Results

The convergences of the performed truss syntheses illustrated in Tables 1–4 (without
buckling) and Tables A1–A4 (with buckling) show that, in all four cases, the identical fi-
nal/optimal result was obtained using both the MTP and STP MINLP strategies. Regarding
the MTP strategy, the optimal result was always obtained after the first two optimization
phases, i.e., after the topology optimization with continuous sizing variables followed by
the discrete sizing optimization at the obtained fixed topology. Up to this point, the MTP
strategy is identical to the original TP strategy [80], which is terminated after the first (and
also final) solution with discrete sizing variables. The simple TP strategy would thus yield
the same final results.
Table 5 illustrates and compares the MINLP iterations required and the CPU times for
different synthesis examples performed by the discussed MINLP strategies.

Table 5. Comparison of optimization statistics for different MINLP strategies.

Synthesis Example
MINLP Strategy
TC10a TC10b TC10c TC10d
TP 60 11 10 7
No. of
MTP 95 16 21 13
MINLP iterations
STP 130 23 21 35
TP 9.657 13.951 1.485 4.324
CPU time (s) MTP 16.650 19.991 3.537 10.889
STP 23.148 21.478 3.344 43.806

The data captured in Table 5 show that the TP strategy required the lowest number
of MINLP iterations as well as the lowest CPU times until the final/optimal result was
attained in all four examples. With regard to the exposed outputs, the TP strategy proved
to be the most effective from among all the compared MINLP strategies. However, as
already mentioned, the TP strategy does not provide any information about the influence
of discrete sizing variables on the possible change in topology. On the other hand, this
Appl. Sci. 2022, 12, 1459 24 of 36

influence is accounted for both in the MTP as well as in the STP MINLP strategy. From
among these two strategies, the MTP strategy proved to be the more convenient one.
It should be noted that the differences between the considered MINLP strategies
become more expressed in cases when the NLP sub-problems become more extensive and
difficult to solve. The CPU times of individual NLP sub-problems are considerably higher
in examples TC10b and TC10d when compared to examples TC10a and TC10c. The former
two examples include shape optimization. With the inclusion of shape variables (nodal
coordinates), not only is the total number of variables higher but the problem is also more
complex as the changing coordinates affect the calculation of internal forces (FE equations).
In example TC10d, buckling constraints are also included. The inter-connection between
buckling constraints and shape variables is additionally present, as the buckling resistance
of a bar depends not only on its cross-sectional area but also on its buckling length. The
CPU time of an individual NLP sub-problem in example TC10d is thus, on average, about
10 times higher than the CPU times of NLP sub-problems in example TC10a. Consequently,
the differences between the applied MINLP strategies become more distinctive.
Although the STP strategy proved to be the least favorable strategy regarding both the
number of MINLP iterations required as well as the CPU times spent, its advantage lies in
the fact that more intermediate solutions are obtained. The first solution with discrete sizing
variables is obtained at the initial topology with all alternative structural elements active.
Considering these solutions, one can gain information about the influence of changeable
topology in the final result with respect to different optimization constraints (with/without
shape optimization and with/without buckling constraints). The optimal discrete sizing
variables and continuous shape variables are therefore presented for both the initial and the
optimal topologies of the four performed syntheses. The results are given in Tables 6 and 7
and graphically presented in Figures 9–12.

Table 6. Optimal results, examples without buckling constraints TC10a (topology optimization) and
TC10b (topology and shape optimization).

Example TC10a Example TC10b

Initial Topology Optimal Topology Initial Topology Optimal Topology
A1,3 (cm2 ) 96.77 96.77 141.94 141.94
A1,4 ( cm2 ) 109.68 109.68 64.51 64.51
A2,3 ( cm2 ) 19.35 19.35 51.61 51.61
A2,4 ( cm2 ) 167.74 180.64 141.94 141.94
A3,4 ( cm2 ) 19.35 19.35 6.45 -
A3,5 ( cm2 ) 96.77 96.77 51.61 51.61
A3,6 ( cm2 ) 6.45 - 51.61 51.61
A4,5 ( cm2 ) 141.94 141.94 6.45 -
A4,6 ( cm2 ) 6.45 - 109.68 109.68
A5,6 ( cm2 ) 6.45 - 96.77 96.77
cx (cm) 827.2361 909.1875
4
cy (cm) 833.5299 801.3623
4
cx (cm) 1566.7881 1580.4308
6
cy (cm) 432.9369 417.6378
6

WEIGHT
19,492.7981 19,266.5406 17,168.1830 16,936.2314
(N)
Appl. Sci. 2022, 12, 1459 25 of 36

Table 7. Optimal results, examples with buckling constraints TC10c (topology optimization) and
TC10d (topology and shape optimization).

Example TC10c Example TC10d

Initial Topology Optimal Topology Initial Topology Optimal Topology
A1,3 (cm2 ) 387.06 387.06 387.06 387.06
A1,4 (cm2 ) 180.64 - 77.42 -
A2,3 (cm2 ) 109.68 109.68 51.61 51.61
A2,4 (cm2 ) 67.74 77.42 109.68 109.68
A3,4 (cm2 ) 6.45 270.94 6.45 154.84
A3,5 (cm2 ) 96.77 270.94 270.94 270.94
A3,6 (cm2 ) 419.25 - 167.74 -
A4,5 (cm2 ) 6.45 96.77 6.45 77.42
A4,6 (cm2 ) 64.51 - 96.77 -
A5,6 (cm2 ) 51.61 - 77.42 -
cx (cm) 303.1833 1311.2365
4
cy (cm) 829.4580 484.7189
4
cx (cm) 1327.1457 -
6
cy (cm) 486.3984 -
6

WEIGHT
41,838.6690 32,204.1317 28,553.4737 26,357.4575
(N)

Figure 9. Optimal solutions, example TC10a (topology optimization, without buckling constraints).
(a) Initial topology; (b) optimal topology.

Figure 10. Optimal solutions, example TC10b (topology and shape optimization, without buckling
constraints). (a) Initial topology; (b) optimal topology.
Appl. Sci. 2022, 12, 1459 26 of 36

Figure 11. Optimal solutions, example TC10c (topology optimization, with buckling constraints).
(a) Initial topology; (b) optimal topology.

Figure 12. Optimal solutions, example TC10d (topology and shape optimization, with buckling
constraints). (a) Initial topology; (b) optimal topology.

Table 6 comprises the results of the optimization without buckling constraints. The
results of example TC10a can be regarded as the solutions of discrete sizing optimization
at the fixed shape and topology (initial topology) and at fixed shape and changeable
topology (optimal topology). Similarly, the results of example TC10b can be regarded as
the solutions of the shape and discrete sizing optimization at the fixed topology (initial
topology) and at the changeable topology (optimal topology). The results presented in
Table 7 can be classified equivalently for the case of optimization with buckling constraints
included. In this way, the achieved improvement in results at different optimization levels
(topology and/or shape optimization) with regard to sizing optimization at fixed shape
and topology (examples TC10a and TC10c at initial topology) can be estimated. The results
of the comparative analysis are presented in Table 8.

Table 8. Improvement in final result (in %) at different optimization levels.

Buckling Optimization Level

Constraints Shape Topology Shape and Topology
Not included 11.93 1.16 13.12
Included 31.75 23.03 37.00

In the case of optimization with stress and displacement constraints only (buckling is
not included), surprisingly little decrease in optimal weight was achieved by the topology
optimization. The shape optimization (at the fixed topology) yielded a considerably better
result. However, when the buckling constraints were included, the topology optimization
led to a considerably higher improvement, while shape optimization still yielded a better
result. As expected in both cases, the best result was obtained by combining the topology
and shape optimizations, yet it is obvious that almost a three times higher improvement in
result (decrease in weight) was achieved when buckling constraints were included when
Appl. Sci. 2022, 12, 1459 27 of 36

compared to the case of the optimization with stress and displacement constraints only
(37.00% vs. 13.12%).
A more detailed analysis of the obtained results leads to a further finding. As has
already been stated in the introductory sections, topology of truss structures was, in the
past, often optimized by allowing very small (nonzero) lower bounds on cross-section
areas of bars. When the cross-section areas of some bars were valued to the defined lower
bounds, these bars were assumed to be redundant and were therefore excluded from the
structure. Consequently, the optimal topology was adopted as the initial topology with the
excluded »lower bound bars«.
When it comes to the results of the proposed MINLP truss synthesis, this assumption
proves to be correct when the buckling constraints are excluded. The obtained optimal
topologies in examples TC10a and TC10b are simply the initial topologies, where the bars
with the lowest discrete cross-section area (6.45 cm2 ) are excluded; see Table 6. This can be
seen very clearly in Figures 9 and 10. On the other hand, when the buckling constraints
are included, the discussed subject turns out to be different. In the optimal topologies of
examples TC10c and TC10d, those bars are active, which, in the initial topology, took the
lowest value. Excluding these bars would therefore lead to a non-optimal topology.
Beside stress, the buckling of compressed bars, is one of the basic conditions that has
to be accounted for in the design process of a truss structure. In the initial topology, the
compressed bars are subjected to buckling constraints and therefore require larger areas of
cross-sections compared to tension bars, which are subjected to stress constraints only. The
method of topology optimization with the elimination the least stressed bars mentioned
above is thus not appropriate for real engineering structures.
Finally, the results of the proposed MINLP truss synthesis are compared to some
results available in literature. The comparison of the optimal solution of example TC10a is
presented in Table 9. The MINLP synthesis yielded the same result as the currently best
solution presented in [49].

Table 9. Comparison of results.

Reference [45] Reference [46] Reference [47] Reference [48] Present Work
A1,3 (cm2 ) 109.68 109.68 96.77 96.77 96.77
A1,4 (cm2 ) 96.77 109.68 109.68 109.68 109.68
A2,3 (cm2 ) 32.26 32.26 19.35 19.35 19.35
A2,4 (cm2 ) 167.74 167.74 167.74 180.64 180.64
A3,4 (cm2 ) 51.61 - 19.35 19.35 19.35
A3,5 (cm2 ) 96.77 77.42 109.68 96.77 96.77
A3,6 (cm2 ) - 32.26 - - -
A4,5 (cm2 ) 141.94 109.68 141.94 141.94 141.94
A4,6 (cm2 ) - 19.35 - - -
A5,6 (cm2 ) - 19.35 - - -
WEIGHT (N) 20,058 19,712 19,266.7886 19,266.5406 19,266.5406

Additionally, all four examples of truss cantilever synthesis were performed using
two commercial MINLP solvers, DICOPT [85] and BARON [86]. The previously discussed
Two-Phase approach and the pre-screening of discrete sizing variables were applied in
all cases.
DIPOCT also represents an implementation of the OA/ER algorithm. GAMS/CPLEX
and GAMS/CONOPT were applied as the solvers for the MILP and NLP sub-problems,
respectively. Two different stopping criteria were used:
1. Stop as soon as the bound defined by the objective of the MILP master problem
is worse than the best NLP solution found (the same criterion as in the previous
examples solved by MIPSYN), and
2. Stop as soon as the NLP sub-problems cease to improve.
Appl. Sci. 2022, 12, 1459 28 of 36

Both criteria yielded the same final results.

By using BARON (Branch-And-Reduce Optimization Navigator), GAMS/CPLEX was
applied as the LP solver, while GAMS/MINOS [87] was applied as the NLP solver. The
maximum allowed CPU time of 10000 s was defined for each optimization phase, and
multiple solutions were allowed.
The results obtained by different MINLP solvers are presented in Table 10. Compared
to the solutions obtained by MIPSYN, DICOPT found the same solution in example TC10a,
while in the remaining examples, worse solutions were found. BARON yielded the same
solutions as MIPSYN in the examples TC10a and TC10c, while the remaining results
(examples TC10b and TC10d) also represented worse solutions. In Table 11, the CPU
times are also presented. For the solvers MISPYN and DICOPT, the CPU time until the
termination of the Two-Phase procedure is shown. By using these two solvers, the CPU
times are comparable, although DICOPT showed its advantages in the examples without
buckling constraints (TC10a and TC10b), while MIPSYN proved to be more efficient in
the remaining two examples. In the case of BARON, on the other hand, the optimization
progress was much slower, and in most cases, the process continued to run until the CPU
time limit of 10,000 s was reached. Therefore, the CPU times shown in Table 11 for the
case of BARON represent the times spent until the best solutions (as shown in Table 10)
were obtained.

Table 10. Results obtained by using different MINLP solvers.

MINLP (NLP/MILP, NLP/LP) Solver

Synthesis
Example DICOPT BARON MIPSYN
(CONOPT/CPLEX) (MINOS/CPLEX) (CONOPT/CPLEX)
TC10a 19,266.5406 19,266.5406 19,266.5406
TC10b 17,078.4080 16,968.0110 16,936.2314
TC10c 32,590.3776 32,204.1317 32,204.1317
TC10d 26,516.4615 26,573.4506 26,357.4575

Table 11. The comparison of the CPU times (in seconds).

MINLP (NLP/MILP, NLP/LP) Solver

Synthesis
Example DICOPT BARON MIPSYN
(CONOPT/CPLEX) (MINOS/CPLEX) (CONOPT/CPLEX)
TC10a 1.01 14.00 9.657
TC10b 1.00 2513.00 13.951
TC10c 3.67 8909.00 1.485
TC10d 16.24 6205.00 4.324

6. Conclusions
The present paper presents the Mixed-Integer Non-linear Programming (MINLP)
approach to truss synthesis. The MINLP approach enables the topology, shape and dis-
crete/standard cross-sectional dimensions to be optimized simultaneously. In the mixed
continuous and discrete type of optimization problem, the continuous variables for con-
tinuous parameters (nodal coordinates, stresses, displacements) as well as the discrete
binary variables for discrete decisions are defined. Binary 0–1 variables define the existence
(1) or non-existence (0) of structural elements and are also subjected to the choice of the
discrete/standard dimensions of cross-sections.
The proposed MINLP truss synthesis is performed through three steps. The first in-
volves the generation of a truss superstructure of different topology and discrete dimension
alternatives; the second presents the development of a special MINLP model formulation
for truss superstructures (MINLP-TS); and the last contains a solution for the defined
MINLP problem. The problems are solved by the Modified OA/ER algorithm.
Appl. Sci. 2022, 12, 1459 29 of 36

Besides several non-linearities and non-convexities involved, the main difficulty of

truss synthesis is that a very high number of discrete variables, particularly those associ-
ated with discrete/standard dimensions, may be included in the optimization. Multi-level
MINLP strategies were thus developed and applied to make the solving of such compre-
hensive optimization problems possible. The main advantage of multi-level strategies is
that the sub-set of binary variables, activated at each level, is considerably smaller than the
full set. The additional reduction in the number of currently active binary variables for
discrete dimensions was attained by the use of a special pre-screening procedure.
Two new types of multi-level strategies were presented, namely the Modified Two-
Phase (MTP) and the Sequential Two-Phase (STP) MINLP strategies. Both strategies were
developed on the basis of the classical Two-Phase (TP) strategy. The motive of developing
new strategies was the fact that the TP strategy does not allow the topology to be changed
after the discrete sizing optimization and therefore does not consider any interaction
between topology and discrete sizing optimization levels. Out of the two new strategies,
the MTP strategy proved to be the more convenient option regarding the MINLP iterations
required and the CPU times spent. Considering the final/optimal result, however, the basic
TP strategy proved also to be a useful tool for obtaining good solutions by spending short
CPU times.
The proposed MINLP synthesis was applied to the simple and well-known numer-
ical example of the ten-bar truss cantilever. The simultaneous topology and discrete
sizing optimization was performed with/without the inclusion of shape optimization
and with/without buckling constraints. The comparison of the obtained results with re-
sults available from literature proved the proposed MINLP synthesis to be a competitive
optimization technique.
Some important conclusions regarding the buckling constraints of compressed bars
also need to be emphasized. Not only do the buckling constraints represent a vital step
of the design process, these constraints also have a crucial influence on the progress of
topology optimization. When a truss is optimized, considering the stress and displacement
constraints only, it is advisable for the low-stressed bars from the initial topology to be
removed from the structure. By introducing buckling constraints, however, the lower
stressed (mostly tensioned) bars from the initial topology are often a better choice for
selection, while the higher stressed (mostly compressed) ones are rejected. Thus, it is not
advisable to remove the low-stressed bars from the initial topology when the buckling
constraints are considered. Rigorous topology such as that proposed in this paper should
be employed to obtain optimal solutions. A considerably greater improvement in results
(reduction in structural weight) was additionally achieved by the introduction of the shape
optimization into the synthesis.

Author Contributions: Conceptualization, S.Š., Z.K. and S.K.; methodology, S.Š., Z.K. and S.K.;
formal analysis, S.Š., Z.K. and S.K.; writing—original draft preparation, S.Š., Z.K. and S.K. All authors
have read and agreed to the published version of the manuscript.
Funding: The authors acknowledge the financial support from the Slovenian Research Agency (grant
numbers P2-0129 and P2-0032).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data presented in this study are available upon request from the
corresponding author.
Conflicts of Interest: The authors declare no conflict of interest.
Appl. Sci. 2022, 12, 1459 30 of 36

Appendix A
Table A1. Convergence of the Modified OA/ER algorithm and the MTP strategy, example TC10c
(topology optimization, with buckling constraints).

Optimization
MINLP Iteration Value of Objective Function No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
1 1. NLP 40,612.5684 - 0.180
2 1. MILP 17,009.8900 8 0.030
2. NLP 31,376.7356 (zt ) 0.148
3 2. MILP 26,993.8600 0.031
T 3. NLP 31,696.7121 0.172
4 3. MILP 28,332.1100 0.050
n=1
4. NLP 35,270.9973 0.121
5 4. MILP 33,960.8800 0.050
5. NLP 35,590.9738 0.109
... ... ... 12
D 10 9. MILP 32,204.1300
10. NLP 32,204.1317 (zOPT ) (*) 0.594
11 10. MILP 32,483.5800 8 0.040
T
11. NLP 31,696.7121 (zt < zOPT ) 0.121
n=2 ... ... ... 14
D 16 15. MILP 32,524.1100
16. NLP 32,524.1082 (zd > zOPT ) (*) 0.832
17 16. MILP 30,427.5500 8 0.050
17. NLP 35,270.9973 (z1 t > zOPT ) 0.148
18 17. MILP 32,423.2100 0.060
18. NLP 35,590.9738 (z2 t > zOPT ) 0.082
19 18. MILP 34,956.8920 0.080
n=3 T 19. NLP 34,630.1695 (z3 t > zOPT ) 0.082
20 19. MILP 35,625.8000 0.081
20. NLP 34,950.1460 (z4 t > zOPT ) 0.117
21 20. MILP 42,980.0600 0.090
21. NLP 35,309.5418 (z5 t > zOPT ) 0.109
 = 3.537
(*)
sum of CPU times for the entire D phases; MINLP phase T: Topology and continuous sizing optimization;
MINLP phase D: Discrete sizing optimization at fixed topology.

Table A2. Convergence of the Modified OA/ER algorithm and the STP strategy, example TC10c
(topology optimization, with buckling constraints).

Optimization
MINLP Iteration Value of Objective Function No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
T 1 1. NLP 40,612.5684 (zt ) - 0.180
... ... ...
n=1
D 4 3. MILP 41,838.6700
4. NLP 41,838.6690 (zOPT ) (*) 0.34
5 4. MILP 17,009.8900 8 0.030
T
5. NLP 31,376.7356 (zt < zOPT ) 0.148
n=2 ... ... ... 12
D 10 9. MILP 32,204.1300
10. NLP 32,204.1317 (zOPT ) (*) 0.594
11 10. MILP 32,483.5800 8 0.040
T
11. NLP 31,696.7121 (zt < zOPT ) 0.121
n=3 ... ... ... 14
D 16 15. MILP 32,524.1100
16. NLP 32,524.1082 (zd > zOPT ) (*) 0.832
Appl. Sci. 2022, 12, 1459 31 of 36

Table A2. Cont.

Optimization
MINLP Iteration Value of Objective Function No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
17 16. MILP 30,427.5500 8 0.050
17. NLP 35,270.9973 (z1 t > zOPT ) 0.148
18 17. MILP 32,423.2100 0.060
18. NLP 35,590.9738 (z2 t > zOPT ) 0.082
19 18. MILP 34,956.8920 0.080
n=4 T 19. NLP 34,630.1695 (z3 t > zOPT ) 0.082
20 19. MILP 35,625.8000 0.081
20. NLP 34,950.1460 (z4 t > zOPT ) 0.117
21 20. MILP 42,980.0600 0.090
21. NLP 35,309.5418 (z5 t > zOPT ) 0.109
 = 3.344
(*)
sum of CPU times for the entire D phases; MINLP phase T: Topology and continuous sizing optimization;
MINLP phase D: Discrete sizing optimization at fixed topology.

Table A3. Convergence of the Modified OA/ER algorithm and the MTP strategy, example TC10d
(topology and shape optimization, with buckling constraints).

Optimization
MINLP Iteration Value of Objective Function No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
1 1. NLP 27,755.0426 - 0.809
2 1. MILP 16,094.2100 8 0.030
2. NLP 31,189.2331 0.691
T 3 2. MILP 22,727.4900 0.040
3. NLP 25,695.6485 (zt ) 1.008
n=1
4 3. MILP 26,774.4300 0.050
4. NLP 38,672.9100 0.723
... ... ... 12
D 7 6. MILP 25,695.6480
7. NLP 26357.4575 (zOPT ) (*) 0.973
8 7. MILP 25,590.0000 8 0.060
8. NLP 31,189.2331 (z1 t > zOPT ) 0.980
9 8. MILP 28,733.3300 0.071
9. NLP 29,436.0921 (z2 t > zOPT ) 0.531
10 9. MILP 31,679.9400 0.070
10. NLP 31,525.3188 (z3 t > zOPT ) 0.672
n=2 T 11 10. MILP 31,679.9440 0.110
11. NLP a 23,004.7324 1.402
12 11. MILP 32,453.3520 0.100
12. NLP 30,543.6905 (z4 t > zOPT ) 1.211
13 12. MILP 36,012.9370 0.100
13. NLP 27,539.0495 (z5 t > zOPT ) 1.258
 = 10.889
(*)sum of CPU times for the entire D phases; a locally infeasible solution; MINLP phase T: Topology and continuous
sizing optimization; MINLP phase D: Discrete sizing optimization at fixed topology.
Appl. Sci. 2022, 12, 1459 32 of 36

Figure A1. Convergence to the optimal topology (initial MINLP phase of the MTP strategy), example
TC10c (topology optimization, with buckling constraints).

Figure A2. Convergence to the optimal topology (initial MINLP phase of the MTP strategy), example
TC10d (topology and shape optimization, with buckling constraints).
Appl. Sci. 2022, 12, 1459 33 of 36

Table A4. Convergence of the Modified OA/ER algorithm and the STP strategy, example TC10d
(topology and shape optimization, with buckling constraints).

Optimization
MINLP Iteration Value of Objective Function No. of Active Bin. Var. CPU Time (s)
Cycle Phase (N)
T 1 1. NLP 27,755.0426 (zt ) - 0.832
... ... ... 20
n=1
D 10 9. MILP 26,036.1220
10. NLP 28,553.4737 (zOPT ) (*) 5.929
11 10. MILP 16,234.9900 8 0.030
11. NLP 31,189.2331 0.621
12 11. MILP 18,417.8610 0.050
T
12. NLP 29,436.0921 0.488
n=2 13 12. MILP 20,849.8900 0.050
13. NLP 25,695.6485 (zt < zOPT ) 0.898
... ... ... 12
D 16 15. MILP 25,695.6480
16. NLP 26,357.4575 (zOPT ) (*) 0.969
17 16. MILP 25,275.5140 8 0.060
17. NLP 38,672.9100 0.551
T
18 17. MILP 26,827.2820 0.090
n=3
18. NLP 25,948.8041 (zt < zOPT ) 12.102
19 18. MILP 31,929.6360 0.090
D
19. NLP 30,400.0000 (zd > zOPT ) 0.988
20 19. MILP 27,143.8800 8 0.100
T
20. NLP 25,794.2192 (zt < zOPT ) 0.891
n=4 ... ... ... 14
D 29 28. MILP 31,141.9100
29. NLP 28,847.4632 (zd > zOPT ) (*) 8.336
30 29. MILP 34,598.6040 8 0.120
30. NLP 31,580.8787 (z1 t > zOPT ) 1.469
31 30. MILP 30,636.5880 0.121
31. NLP 29,643.5685 (z2 t > zOPT ) 3.168
32 31. MILP 34,500.2680 0.170
32. NLP a 19,762.7370 0.309
n=5 T 33 32. MILP 29,470.9030 0.140
33. NLP 29,523.1464 (z3 t > zOPT ) 2.520
34 33. MILP 23,966.5330 0.100
34. NLP 28,227.8465 (z4 t > zOPT ) 1.199
35 34. MILP 30,538.3790 0.180
35. NLP 29,673.0156 (z5 t > zOPT ) 1.235
 = 43.806
(*)
sum of CPU times for the entire D phases; MINLP phase T: Topology and continuous sizing optimization;
MINLP phase D: Discrete sizing optimization at fixed topology.

References
1. Maxwell, J.C. On Reciprocal Figures, Frames and Diagrams of Forces. Trans. R. Soc. Edinb. 1869, 26, 1–40. [CrossRef]
2. Michell, A.G.M. The Limits of Economy in Frame Structures. Philos. Mag. 1904, 8, 589–597.
3. Schmit, L.A. Structural Design by Systematic Synthesis. In Proceedings of the 2nd Conference on Electronic Computations; ASCE:
New York, NY, USA, 1960; pp. 105–122.
4. Kravanja, Z.; Grossmann, I.E. New developments and capabilities in PROSYN—An automated topology and parameter synthe-
sizer. Comput. Chem. Eng. 1994, 18, 1097–1114. [CrossRef]
5. Dorn, W.S.; Gomory, R.E.; Greenberg, H. Automatic Design of Optimal Structures. J. Mécanique 1964, 3, 25–52.
6. Sved, G.; Ginos, Z. Structural Optimization Under Multiple Loading. Int. J. Mech. Sci. 1968, 10, 803–805. [CrossRef]
7. Sheu, C.Y.; Schmit, L.A. Minimum Weight Design of Elastic Redundant Trusses Under Multiple Static Loading Conditions. AIAA
J. 1972, 10, 155–162. [CrossRef]
8. Kirsch, U.; Taye, S. On Optimal Topology of Grillage Structures. Eng. Comput. 1986, 1, 229–243. [CrossRef]
Appl. Sci. 2022, 12, 1459 34 of 36

9. Kirsch, U.; Topping, B.H.V. Minimum Weight Designs of Structural Topologies. J. Struct. Eng. 1992, 118, 1770–1785. [CrossRef]
10. Sankaranarayanan, S.; Haftka, R.T.; Kapania, R.K. Truss topology optimization with simultaneous analysis and design. AIAA J.
1994, 32, 420–424. [CrossRef]
11. Achtziger, W. On simultaneous optimization of truss geometry and topology. Struct. Multidiscip. Optim. 2007, 33, 285–304.
[CrossRef]
12. Goldberg, D.E. Genetic Algorithms in Search, Optimization and Machine Learning; Addison-Wesley: Reading, MA, USA, 1989.
13. Hajela, P.; Lee, E. Genetic Algorithms in Truss Topology Optimization. Int. J. Solids Struct. 1995, 32, 3341–3357. [CrossRef]
14. Rajan, S.D. Sizing, Shape and Topology Design Optimization of Trusses using Genetic Algorithm. J. Struct. Eng. 1995,
121, 1480–1487. [CrossRef]
15. Rajeev, S.; Krishnamoorthy, C.S. Genetic Algorithms Based Methodologies for Design Optimization of Trusses. J. Struct. Eng.
1997, 123, 350–358. [CrossRef]
16. Kawamura, H.; Ohmori, H.; Kito, N. Truss Topology Optimization by a Modified Genetic Algorithm. Struct. Multidiscip. Optim.
2002, 23, 467–472. [CrossRef]
17. Cagan, J.; Mitchell, W.J. Optimally Directed Shape Generation by Shape Annealing. Environ. Plan. 1993, 20, 5–12. [CrossRef]
18. Reddy, G.; Cagan, J. An Improved Shape Annealing Algorithm for Truss Topology Optimization. J. Mech. Des. 1995, 117, 315–321.
[CrossRef]
19. Hasancebi, O.; Erbatur, F. Layout Optimisation of Trusses using Simulated Annealing. Adv. Eng. Softw. 2002, 33, 681–696.
[CrossRef]
20. Sandgren, E. Multiple-objective, shape optimal design via genetic optimization. In Computer Aided Optimium Design of Structures
IV; Hernandes, S., El-Sayed, M., Brebbia, C.A., Eds.; Computational mechanics publications: Southampton, UK; Boston, MA,
USA, 1995; pp. 3–10.
21. Bennage, W.A.; Dhingra, A.K. Optimization of Truss Topology using Tabu Search. Int. J. Numer. Methods Eng. 1995, 38, 4035–4052.
[CrossRef]
22. Bendsøe, M.P.; Kikuchi, N. Generating Optimal Topologies in Structural Design using a Homogenization Method. Comput.
Methods Appl. Mech. Eng. 1988, 71, 197–224. [CrossRef]
23. Suzuki, K.; Kikuchi, N. A Homogenization Method for Shape and Topology Optimization. Comput. Methods Appl. Mech. Eng.
1991, 93, 291–318. [CrossRef]
24. Diaz, A.R.; Belding, B. On Optimum Truss Layout by a Homogenization Method. J. Mech. Des. 1993, 115, 367–373. [CrossRef]
25. Yildiz, A.R.; Öztürk, N.; Kaya, N.; Öztürk, F. Integrated Optimal Topology Design and Shape Optimization using Neural
Networks. Struct. Multidiscip. Optim. 2003, 25, 251–260. [CrossRef]
26. Toakley, R. Optimum Design using Available Sections. ASME J. Struct. Div. 1968, 94, 1219–1241. [CrossRef]
27. Templeman, A.B.; Yates, D.F. A Linear Programming Approach to Discrete Optimum Design of Trusses. In Optimization Methods
in Structural Design; Eschenauer, H., Olhoff, N., Eds.; BI Wissenschaftsverlag: Mannheim, Germany, 1983.
28. Zhou, D.M. An Improved Templeman’s Algorithm for Optimum Design of Trusses with Discrete Member Sizes. Eng. Optim.
1986, 9, 303–312.
29. John, K.V.; Ramakrishnan, C.V. Minimum Weight Design of Trusses using Improved Move Limit of Sequential Linear Program-
ming. Int. J. Comput. Struct. 1987, 27, 583–591. [CrossRef]
30. Bremicker, M.; Papalambros, P.Y.; Loh, H.T. Solution of Mixed-Discrete Structural Optimization Problems with a new Sequential
Linearization Algorithm. Comput. Struct. 1990, 37, 451–461. [CrossRef]
31. Salajegeh, E.; Vaderplaats, G.N. Optimum Design of Trusses with Discrete Sizing and Shape Variables. Struct. Optim. 1993,
6, 79–85. [CrossRef]
32. Rajeev, S.; Krishnamoorthy, C.S. Discrete Optimization of Structures using Genetic Algorithms. J. Struct. Eng. 1992, 118, 1233–1250.
[CrossRef]
33. Lin, C.Y.; Hajela, P. Genetic Algorithms in Optimization Problems with Discrete and Integer Design Variables. Eng. Optim. 1992,
19, 309–327. [CrossRef]
34. Erbatur, F.; Hasancebi, O.; Tütüncü, I.; Kilic, H. Optimal Design of Planar and Space Trusses with Genetic Algorithms. Comput.
Struct. 2000, 75, 209–224. [CrossRef]
35. May, S.A.; Balling, R.J. A Filtered Simulated Annealing Strategy for 3D Optimization of Steel Frameworks. Struct. Optim. 1992,
4, 142–148. [CrossRef]
36. Pantelides, C.P.; Tzan, S.R. Optimal Design of Dynamically Constrained Structures. Comput. Struct. 1997, 62, 141–149. [CrossRef]
37. Cai, J.B.; Thiereut, G. Discrete Optimization of Structures using an Improved Penalty Function Method. Eng. Optim. 1993,
21, 293–306. [CrossRef]
38. Shih, C.J. Fuzzy and Improved Penalty Approaches for Multiobjective Mixed-Discrete Optimization in Structural Systems.
Comput. Struct. 1997, 6, 559–565. [CrossRef]
39. Shih, C.J.; Yang, Y.C. Generalized Hopfield Network Based Structural Optimization using Sequential Unconstrained Minimization
Technique with Additional Penalty Strategy. Adv. Eng. Softw. 2002, 33, 721–729. [CrossRef]
40. Jivotovski, G.A. Gradient Based Heuristic Algorithm and its Application to Discrete Optimization of Bar Structures. Struct.
Multidiscip. Optim. 2000, 19, 237–248. [CrossRef]
Appl. Sci. 2022, 12, 1459 35 of 36

41. Tong, W.H.; Liu, G.R. An Optimization Procedure for Truss Structures with Discrete Design Variables and Dynamic Constraints.
Comput. Struct. 2001, 79, 155–162. [CrossRef]
42. Guerlement, G.; Targowski, R.; Gutkowski, W.; Zawidzka, J.; Zawidzki, J. Discrete Minimum Weight Design of Steel Structures
using EC3 Code. Strucural Multidiscip. Optim. 2001, 22, 322–327. [CrossRef]
43. Eurocode 3: Design of Steel Structures; European Comitee for Standardization: Bruxelles, Belgium, 2005.
44. Bollapragada, S.; Ghattas, O.; Hooker, J.N. Optimal Design of Truss Structures by Logic-based Branch and Cut. Oper. Res. 2001,
49, 42–51. [CrossRef]
45. Ohsaki, M. Random Search Method based on Exact Reanalysis for Topology Optimization of Trusses with Discrete Cross-sectional
Areas. Comput. Struct. 2001, 79, 673–679. [CrossRef]
46. Mela, K. Resolving issues with member buckling in truss topology optimization using a mixed variable approach. Struct Multidisc
Optim 2014, 50, 1037–1049. [CrossRef]
47. Wang, Y.; Sun, H. The Topology Optimization of Structure with Discrete Variables under Multiload Case and Multiconstraint.
Acta Mech. Sin. 1995, 27, 365–369.
48. Chai, S.; Shi, L.S.; Sun, H.C. An Application of Relative Difference Quotient Algorithm to Topology Optimization of Truss
Structures with Discrete Variables. Struct. Optim. 1999, 18, 48–55. [CrossRef]
49. Kaveh, A.; Kalatjari, V. Topology Optimization of Trusses using Genetic Algorithm, Force Method and Graph Theory. Int. J.
Numer. Methods Eng. 2003, 58, 771–791. [CrossRef]
50. Kaveh, A.; Shahrouzi, M. Simultaneous topology and size optimization of structures by genetic algorithm using minimal length
chromosome. Eng. Comput. 2006, 23, 644–674. [CrossRef]
51. Kaveh, A.; Shojaee, S. Optimal design of skeletal structures using ant colony approach. Int. J. Numer. Methods Eng. 2007,
70, 563–581. [CrossRef]
52. Kaveh, A.; Shahrouzi, M. Optimal structural design family by genetic search and ant colony approach. Eng. Comput. 2008,
25, 268–288. [CrossRef]
53. Chen, S.Y.; Shui, X.F.; Li, D.F.; Huang, H. Improved genetic algorithm with two-level approximation for truss topology optimiza-
tion. Math. Probl. Eng. 2015, 2015, 521482.
54. Kanno, J. Global optimization of trusses with constraints on number of different cross-sections: A mixed-integer second-order
cone programming approach. Comput Optim Appl 2016, 63, 203–236. [CrossRef]
55. Mortazavi, A.; Togan, V. Simultaneous size, shape, and topology optimization of truss structures using integrated particle swarm
optimizer. Struct Multidisc Optim 2016, 54, 715–736. [CrossRef]
56. Savsani, V.J.; Tejani, G.G.; Patel, V.K.; Savsani, P. Modified meta-heuristics using random mutation for truss topology optimization
with static and dynamic constraints. J. Comput. Des. Eng. 2017, 4, 106–130. [CrossRef]
57. Degertekin, S.O.; Lamberti, L.; Ugur, I.B. Discrete sizing/layout/topology optimization of truss structures with an advanced Jaya
algorithm. Appl. Soft Comput. 2019, 79, 363–390. [CrossRef]
58. Shahabsafa, M.; Fakhimi, R.; Lei, W.; He, S.; Martins, J.R.R.A.; Terlaky, T.; Zuluaga, L.F. Truss topology design and sizing
optimization with guaranteedkinematic stability. Struct. Multidisc. Optim. 2021, 63, 21–38. [CrossRef]
59. Kravanja, S.; Kravanja, Z.; Bedenik, B.S. The MINLP approach to structural synthesis, Part I: A general view on simultaneous
topology and parameter optimization. Int. J. Numer. Methods Eng. 1998, 43, 263–292. [CrossRef]
60. Kravanja, S.; Kravanja, Z.; Bedenik, B.S. The MINLP approach to structural synthesis, Part II: Simultaneous topology, parameter
and standard dimension optimization by the use of the linked two-phase strategy. Int. J. Numer. Methods Eng. 1998, 43, 293–328.
[CrossRef]
61. Kravanja, S.; Kravanja, Z.; Bedenik, B.S. The MINLP approach to structural synthesis, Part III: Synthesis of roller and sliding
hydraulic steel gate structures. Int. J. Numer. Methods Eng. 1998, 43, 329–364. [CrossRef]
62. Kravanja, S. Optimization of the Sultartangi sliding gates in Iceland. Int. J. Hydropower Dams 2002, 9, 42–45.
63. Kravanja, S.; Turkalj, G.; Šilih, S.; Žula, T. Optimal design of single-story steel building structures based on parametric MINLP
optimization. J. Constr. Steel Res. 2013, 81, 86–103. [CrossRef]
64. Kravanja, S.; Klanšek, U.; Žula, T. Mass, Direct Cost and Energy Life-Cycle Cost Optimization of Steel-Concrete Composite Floor
Structures. Appl. Sci. 2021, 11, 10316. [CrossRef]
65. Šilih, S.; Kravanja, S.; Bedenik, B.S. Finite Elements in Civil Engineering Applications: Proceedings of the Third Diana World Conference,
Tokyo, Japan, 9–11 October 2002; Hendriks, M.A.N., Rots, J.G., Eds.; Swets & Zeitlinger: Lisse, The Netherlands, 2002; pp. 369–373.
66. Kravanja, S.; Šilih, S. Optimization based comparison between composite I beams and composite trusses. J. Constr. Steel Res. 2003,
59, 609–625. [CrossRef]
67. Šilih, S.; Premrov, M.; Kravanja, S. Optimum design of plane timber trusses considering joint flexibility. Eng. Struct. 2005,
27, 145–154. [CrossRef]
68. Beale, E.M.L. Integer programming. In The State of the Art in Numerical Analysis; Jacobs, D., Ed.; Academic Press: London, UK,
1977; pp. 409–448.
69. Gupta, O.K.; Ravindran, A. A nonlinear mixed intefer programming and discrete optimization. In Progress in Engineering
Optimization; Mayne, R.W., Ragsdell, K.M., Eds.; ASME: New York, NY, USA, 1984; pp. 295–520.
70. Quesada, I.; Grossmann, I.E. An LP/NLP based branch and bound algorithm for convex MINLP optimization problems. Comput.
Chem. Eng. 1992, 16, 937–947. [CrossRef]
Appl. Sci. 2022, 12, 1459 36 of 36

71. Westerlund, T.; Pettersson, F. An extended cutting plane method for solving convex MINLP problems. Comput. Chem. Eng. 1995,
19, 131–136. [CrossRef]
72. Benders, J.F. Partitioning procedures for solving mixed-variables programming problems. Numer. Math. 1962, 4, 238–252.
[CrossRef]
73. Geoffrion, A.M. Generalized Benders decomposition. J. Optim. Theory Appl. 1972, 10, 237–260. [CrossRef]
74. Duran, M.A.; Grossmann, I.E. An Outer-Approximation Algorithm for a Class of Mixed-Integer Nonlinear Programms. Math.
Program. 1986, 36, 307–339. [CrossRef]
75. Mawengkang, H.; Murtagh, B.A. Solving nonlinear integer programs with large-scale optimization software. Annu. Opereatins
Res. 1986, 5, 425–437. [CrossRef]
76. Kocis, G.R.; Grossmann, I.E. Relaxation Strategy for the Structural Optimization of Process Flowsheets. Inustrial Eng. Chem. Res.
1987, 26, 1869–1880. [CrossRef]
77. Kravanja, S.; Bedenik, B.S.; Kravanja, Z. MINLP optimization of mechanical structures. In Structural Optimization 93, The World
Congress on Optimal Design of Structural Systems, Volume 1; Herskovits, J., Ed.; Federal University of Rio de Janiero: Rio de Janeiro,
Brasil, 1993; pp. 21–28.
78. Kravanja, S.; Soršak, A.; Kravanja, Z. Efficient multilevel MINLP strategies for solving large combinatorial problems in engineering.
Optim. Eng. 2003, 4, 97–151. [CrossRef]
79. Kravanja, S.; Šilih, S.; Kravanja, Z. The multilevel MINLP optimization approach to structural synthesis: The simultaneous
topology, material, standard and rounded dimension optimization. Intern. J. Adv. Eng. Softw. 2005, 36, 568–583. [CrossRef]
80. Kravanja, S.; Kravanja, Z.; Bedenik, B.S.; Faith, S. Simultaneous topology and parameter optimization of mechanical struc-
tures. In Numerical Methods in Engineering 92, First European Conference on Numerical Methods in Engineering; Hirsh, C.,
Zienkiewicz, O.C., Eds.; Elsevier: Amsterdam, The Netherlands, 1992; pp. 487–495.
81. Brooke, A.; Kendrick, D.; Meeraus, A. GAMS (General Algebraic Modelling System), a User’s Guide; The Scientific Press: Redwood
City, CA, USA, 1988.
82. Kravanja, Z. Challenges in sustainable integrated process synthesis and the capabilities of an MINLP process synthesizer MipSyn.
Comput. Chem. Eng 2010, 34, 1831–1848. [CrossRef]
83. Drudd, A.S. CONOPT–A Large-Scale GRG Code. ORSA J. Comput. 1994, 6, 207–216. [CrossRef]
84. GAMS/CPLEX User Notes; ILOG Inc.: Geneva, Switzerland. Available online: https://www.gams.com/docs/pdf/cplexman.PDF
(accessed on 2 December 2021).
85. Viswanathan, J.; Grossmann, I.E. A combined Penalty Function and Outer Approximation Method for MINLP Optimization.
Comput. Chem. Eng. 1990, 14, 769–782. [CrossRef]
86. Sahnidis, N. BARON: A general purpose global optimization software package. J. Glob. Optim. 1996, 8, 201–205. [CrossRef]
87. Murtagh, B.A.; Saunders, M.A. MINOS 5.1 User’s Guide, Report SOL 83-20R; Stanford University, Department of Operations
Research: Stanford, CA, USA, 1987.