Anda di halaman 1dari 44

Accepted Manuscript

A hybrid algorithm coupling genetic programming and Nelder-Mead


for topology and size optimization of trusses with static and dynamic
constraints

Hirad Assimi, Ali Jamali

PII: S0957-4174(17)30794-7
DOI: 10.1016/j.eswa.2017.11.035
Reference: ESWA 11678

To appear in: Expert Systems With Applications

Received date: 25 September 2017


Revised date: 14 November 2017
Accepted date: 15 November 2017

Please cite this article as: Hirad Assimi, Ali Jamali, A hybrid algorithm coupling genetic programming
and Nelder-Mead for topology and size optimization of trusses with static and dynamic constraints,
Expert Systems With Applications (2017), doi: 10.1016/j.eswa.2017.11.035

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT

Highlights
• Genetic programming used for truss optimization with static and dy-
namic constraints

• Nelder-Mead used to improve the convergence of the proposed algo-

T
rithm

IP
• The proposed algorithm performed on discrete sizing optimization of
trusses

CR
• The proposed approach outperformed other reported methods in most
of the cases

US
AN
M
ED
PT
CE
AC

1
ACCEPTED MANUSCRIPT

A hybrid algorithm coupling genetic programming and


Nelder-Mead for topology and size optimization of
trusses with static and dynamic constraints

T
Hirad Assimia , Ali Jamalia,∗

IP
a
Faculty of Mechanical Engineering, University of Guilan, P.O. Box: 3756, Rasht, Iran

CR
Abstract

US
Truss optimization aims to provide the lightest truss to gain the maximum
benefit out of available resources. Truss optimization may subject to static
and dynamic constraints. Static constraints include structural kinematic
AN
stability, maximum allowable stress in truss members, maximum allowable
deflection in the truss nodes and critical buckling load. However, dynamic
constraints impose limits on the natural frequency of the desired truss to
avoid the destructive resonance phenomenon. Taking both static and dy-
M

namic constraints into account may lead to growth in the search space but
dwindling its feasible region; the search space becomes very non-convex and
may subterfuge the solver to trap in a local optimum. Another design con-
ED

sideration may include fabricational constraints to present design variables


from a set of available cross-sections to satisfy the design codes. This paper
proposes a hybrid genetic programming algorithm to deal with the barriers
of this complex problem. It looks for the optimum connectivity table (among
PT

the truss nodes) and optimal cross-sectional areas for its members subject to
design constraints. It also benefits from a Nelder-Mead local search operator
to improve the competence and true convergence of the algorithm. Our algo-
CE

rithm has been applied to some numerical examples considering both types
of continuous and discrete design variables; It proved its efficiency to find
better solutions (lighter trusses) in comparison with other methods in the
AC

literature for most of the cases.


Corresponding author
Email addresses: assimi@msc.guilan.ac.ir (Hirad Assimi),
ali.jamali@guilan.ac.ir (Ali Jamali)

Preprint submitted to Expert Systems with Applications November 15, 2017


ACCEPTED MANUSCRIPT

Keywords: Topology optimization, Genetic programming, Nelder-Mead,


Static and dynamic constraints, Truss

1. Introduction

T
Truss optimization problems have been attracted numerous researchers

IP
in late decades to develop more efficient methods. They aim to provide a
preliminary and quick design for further analysis to implement in the real-

CR
world. Truss optimization classifies into three categories: topology optimiza-
tion looks for the optimal connectivity among the truss nodes; size opti-
mization assigns the optimal cross-section areas for the members and shape
optimization seeks the optimum geometric coordinates for the truss nodes in
the design space (Deb, 2001).
US
Truss optimization problems commonly aim to minimize its structural
AN
weight subject to some design constraints. Numerous studies have been
done to optimize truss structures subject to static constraints such as max-
imum stress and deflection in the truss in the last three decades (Yang and
Soh, 1998, Deb, 2001, Li et al., 2007, Mortazavi and Toğan, 2017). Some
M

studies recently have investigated truss optimization subject to natural fre-


quency constraints (Bellagamba and Yang, 1981, Grandhi and Venkayya,
1988, Sedaghati et al., 2002, Wang et al., 2004, Gomes, 2011, Miguel and
ED

Miguel, 2012, Kaveh and Mahdavi, 2015, Kaveh and Ilchi Ghazaan, 2015,
Farshchin et al., 2016a,b). Changing the configuration and size of a truss af-
fect its vibration modes and natural frequencies. Many real-world structures
PT

are subject to external excitations which become dangerous if it leads to the


resonance phenomenon. Dynamic constraints such as limits on the natural
frequencies play a key role to avoid this destructive state of the structure.
CE

Some studies consider discrete cross-section areas for truss optimization


(Li et al., 2009, Kaveh and Mahdavi, 2014, Ho-Huu et al., 2016, Cheng, 2016,
Stolpe, 2016). They choose discrete design variables from a pre-defined set
AC

of available cross-sections to incorporate fabricational limits and satisfy the


provisions in design codes. Discrete design variables lead to discontinuity of
the search space, and it produces obstacles for the problem solver to converge
to global optimum solution (Wu and Chow, 1995).

3
ACCEPTED MANUSCRIPT

Studies which incorporate both static and dynamic constraints are scarce.
Xu et al. applied a parabolic interpolation method called 1-D search method
for this purpose (Xu et al., 2003). They proposed an imaginary bar approach
and topology group method to handle this problem. Kaveh et al. (Kaveh
and Zolghadr, 2013) utilized charged search system algorithm and Savsani et

T
al. applied teaching learning based optimization algorithm and its modified

IP
version for truss optimization subject to both static and dynamic constraints
(Savsani et al., 2016). They later proposed various algorithms utilizing the
random mutation operator for this purpose and incorporated discrete design

CR
variables (Savsani et al., 2017).
Considering both types of constraints turns the search space into a highly
nonlinear and non-convex search space that increases the probability of be-

US
ing swamped in local optimums of the search space. The nature of frequency
constraints are implicit with respect to the design variables and switching
the vibration modes can bring up a challenge to reach the convergence by
the solver (Grandhi, 1993). As the search space grows larger and becomes
AN
more complex, it requires a competent method and a global optimization al-
gorithm to deal with this problem. Evolutionary algorithms have been used
effectively in solving truss optimization problems (Hare et al., 2013). How-
M

ever, evolutionary algorithms are computationally expensive (Kelner et al.,


2008), but it can be considered trifling due to recent developments in com-
puting systems and parallel processing.
ED

Genetic programming is an efficient evolutionary algorithm that proved


itself in solving various problems in majors such as control engineering, mod-
eling, artificial intelligence, neural networks (Poli and Koza, 2014, Jamali
PT

et al., 2014, Gholaminezhad et al., 2016, Jamali et al., 2016). Genetic Pro-
gramming has been also employed for optimizing and evolving some opti-
mization methods (Koohestani and Poli, 2014, Fajfar et al., 2017). Genetic
CE

programming employs a blind-randomly searched population, it improves its


competence by its genetic operators through limited iterations and designates
the best-found solution as the final optimum solution.
AC

Genetic programming has been applied to some structural optimization


studies for different purposes: Soh and Yang employed it for topology, sizing
and shape optimization of trusses (Soh and Yang, 2000, Yang and Soh, 2000,
2002). Zheng et al. developed a linear genetic programming and applied it
to topology, shape and size optimization of truss structures. Fenton et al.

4
ACCEPTED MANUSCRIPT

(2014) applied it to topology and size optimization of trusses and recently


employed it to handle structural optimization with ill-defined boundary con-
ditions (Fenton et al., 2016). Assimi et al. (2017) proposed a structural opti-
mization genetic programming (SOGP) for sizing and topology optimization
of trusses. SOGP benefits from an alteration operator to improve the true

T
convergence of the algorithm to the optimum solution. This alteration op-

IP
erator proposed to overcome the drawback of stationary elements in genetic
programming individuals when no mathematical operator is present in its
function set.

CR
It is noticeable that all studies as mentioned earlier, employ GP for opti-
mization of trusses subject to static constraints. In this paper, we investigate
the efficiency of GP in solving the highly nonlinear and very nonconvex prob-

US
lem of truss optimization with static and dynamic constraints. This paper
proposes a hybrid genetic programming algorithm using a Nelder-Mead local
search operator besides the conventional genetic operators (such as crossover
and mutation). We also consider two different types of design variables as
AN
continuous and discrete design variables to incorporate fabricational con-
straints.
M

Section 2 states the problem and its governing equations; it also expresses
the design considerations and constraints. Section 3 provides the flowchart
of the proposed algorithm and describes its features in details. Section 4
ED

presents some numerical examples to investigate the efficiency of the pro-


posed algorithm and discusses the obtained results in comparison with other
methods reported in the literature.
PT

2. Formulation of the problem


This section describes the problem statement and its governing equa-
CE

tions. Ground structure method proposes an efficient way to approach truss


optimization. It presents the worst structure or the heaviest structure and
contains all possible connections among the nodes (Dorn, 1964); it expresses
AC

the boundary conditions of the optimization problem such as geometric co-


ordination of nodes; forces exerted on the truss, size of the design space and
other design considerations. Truss optimization methods look for the opti-
mum layout (lightest truss) out of the ground structure.

5
ACCEPTED MANUSCRIPT

Nodes in the ground structure categorize into essential and optional types.
Essential nodes include truss nodes that must be present in the optimum de-
sign; these nodes either support the structure, carry the forces or lumped
masses. In contrast, optional nodes are not mandatory to exist in the final
optimum design, and they may get incorporated to prevent the truss turning

T
into a mechanism or better distribution of the stresses through the members.

IP
Sizing and topology truss optimization problem aims to minimize the cost
of a structure by finding the optimal connectivity table among its truss nodes

CR
while it preserves or eliminates truss members from the ground structure. It
intends to distribute the stress throughout the present members and simulta-
neously, it seeks the optimum size of the cross-section areas for each member
to reduce the cost of the truss.
US
The problem of sizing and topology of truss optimization can be stated
as follows (Xu et al., 2003),
AN

m
X n
X
Minimize f (A) = ρi Ai li + cj i = 1, 2, ..., m, j = 1, 2, ..., n
M

i=1 j=1

Subjected to G1 ≡ Structure is suitable


G2 ≡ Structure is kinematically stable
ED

G3 ≡ σi ≤ σimax i = 1, 2, .., m
cr
G4 ≡ σi ≤ σi i = 1, 2, .., m for compressive members
max
G5 ≡ δj ≤ δj j = 1, 2, .., n
PT

G6 ≡ fk ≤ fkmax for some natural frequencies k


G7 ≡ fl ≥ flmax for some natural frequencies l
CE

G8 ≡ Ai ∈ [Amin
i , Amax
i ] i = 1, 2, ...m
(1)
AC

f (A) denotes the structural cost of the truss, and it sums the structural
weight of the truss members with the cost of its nodes, and it depends di-
rectly on the cross-section areas of the members. It should be noted that the
cost functions excludes the lumped masses on the truss nodes. Ai , li and ρi ,
cj denote the cross-section area, length, material density of the ith member

6
ACCEPTED MANUSCRIPT

and mass of the jth node, respectively. Design variables include the cross-
sectional areas of the truss members. Ai denotes the size of the cross-section
area of ith member. Design variables can be continuous or discrete. Con-
tinuous design variables incorporate a range of numbers from a lower limit
(Amin ) to an upper bound (Amax ); but, discrete design variables consider a

T
i i
set of available cross-section areas to choose from; it aims to satisfy existing

IP
provisions in design codes and to incorporate fabricational constraints.

σi , σicr and σimax signify the stress in ith member,the critical buckling

CR
load of the ith member and maximum allowable stress in the truss defined
by the problem, respectively. fk , fkmax denote the kth natural frequency and
its maximum limit; and fl and flmin denote the lth natural frequency and its

of truss nodes, respectively. Amin


i US
minimum limit. m, n indicate the number of truss members, and the number
, Amax
i denote the lower and upper limit of
the cross-sectional area of the truss members.
When a truss member subjects to compressive stress, it may fail if it exceeds
AN
its critical buckling load which is calculated as follows.
ki Ai E
σicr = (2)
li2
M

where E denotes the modulus of elasticity, and ki is the buckling coeffi-


ED

cient.

The constraints from the Equation 1 are inspired from (Deb, 2001) and
PT

are handled using a penalty approach. If a truss violates a constraint, a cost


of penalty is assigned to its structural cost (f (A)) as follows:
CE

F (A) = f (A) + P (3)


AC

7
ACCEPTED MANUSCRIPT

where P is defined:

9
10


if G1 is violated,



 108 if G2 is violated in first step,

107 if G2 is violated in the second step,

T
P = Pm

 105
( | hG3i i | Otherwise.

 Pm i=1 P

IP

 + i=1 | hG4i i | + nj=1 | hG5j i |

 P P
+ k=1 | hG6k i | + l=1 | hG7l i |)

CR
(4)
The following defines each of the design constraints:
1. constraint G1: This constraint tests the suitability of a truss. It

US
checks if the truss includes all essential nodes of the ground structure.
If a truss violates this constraint, it assigns a huge penalty to the truss
structural cost according to Equation 4.
2. constraint G2: If a truss loses stability can turn into a mechanism and
AN
it is undesirable for the designer. One method to verify the stability of
a truss is checking the positive definiteness of its stiffness matrix. This
process is computationally expensive, but Deb and Gulati (Deb, 2001)
M

suggested a two stage approach to handle this issue. In the first stage,
Grubler criterion evaluates the degrees of freedom in the truss, and if it
returns a positive value, it ensures that the structure is a mechanism;
ED

hence, it assigns a significant penalty to the truss structural cost 4.


However, if it returns a non-positive value, it undergoes the second
stage, it checks the positive-definiteness of the stiffness matrix, and if
it violates this constraint, it assigns a big penalty value to the truss
PT

structural cost.
3. constraint G3-G7: These constraints check static and dynamic char-
acteristics of a truss. Constraint G3 checks if stress in a truss member
CE

exceeds the maximum allowable stress. Constraint G4 tests the truss


for failed members because of buckling. Constraint G5 inspects if a
node deflects more than the allowable nodal displacement. Constraint
AC

G6 and G7 check if a truss abides by the natural frequencies limits


required by the design considerations.
if a truss violates any of these limitations, it assigns a cost of a penalty
to the truss structural cost based on a bracket operator penalty (which
is denoted by | hi |) (Deb and Kalyanmoy, 2001) According to the Equa-
tion 4.

8
ACCEPTED MANUSCRIPT

4. Constraint G8: This constraint checks if the design variables are


selected from the allowable range. GP automatically satisfies this con-
straint because it chooses its design variables from the specified range
or a set of available design variables for continuous and discrete design
variables, respectively.

T
It’s noticeable that if a truss violates each of the constraints G1-G2, it

IP
gets penalized and no further calculation is done.

3. Hybrid genetic programming

CR
Figure 1 illustrates the proposed hybrid genetic programming (HGP) al-
gorithm and this section expresses the features of the algorithm.

US
1. Ground strucutre: In the first step, the designer inputs the known
data of the design space to the algorithm. These data include the
AN
geometric coordination of nodes, the possible connectivity table among
nodes, load conditions, lumped masses and specifying the essential and
optional nodes.
2. Initial population: In this step, Genetic programming generates an
M

initial population composed of trees. each tree represents a computer


program which resembles a potential solution to the optimization prob-
lem. These trees compose of branches and leaves which we call them
ED

functions and terminals, respectively. Functions connect inner nodes of


the tree and branches represent the end-point of a subtree. Function
and terminal sets play a key role to obtain potential solutions congru-
PT

ent with the problem. In simple genetic programming, functions may


include mathematical operators or custom functions; the terminal set
contain variables, constants. In order to employ genetic programming
CE

for truss optimization problem, the function and terminal sets are con-
sidered as follows to produce the candidate solutions.
AC

Function set = {A1 , A2 , ..., Am }


Terminal set = {N1 (x1 , y1 , z1 ), N2 (x2 , y2 , z2 ), ..., Nn (xn , yn , zn )}

where Ai denotes the size of cross-sectional area of the ith truss member
and Nj (xj , yj , zj ) indicates the geometric coordinates of the jth node
in the design space. Figure 2 (a) depicts a sample tree generated using

9
ACCEPTED MANUSCRIPT

T
2.5 m 2.5 m 2.5 m
Tree #1

IP
F1
Truss #1

1.5 m
F2

Tree #2

2.5 m
Truss #2
Input

CR
Y F2 F1 User
X
Truss #PS Tree #PS Added mass

Decoding Initial population Ground structure

[K][δ]=[F]
US Parents Offsprings
AN
? kg [K][M]
Weight calculation Static and modal analysis Crossover operator
M

Selected tree

No Termination
criterion
ED

satisfied? Offspring
Generated
Genetic programming individual subtree

Nelder-mead operator Mutation operator

Yes
PT
CE

Optimum design
AC

Figure 1: The flowchart of the proposed hybrid genetic programming

10
ACCEPTED MANUSCRIPT

A1

A5 A6

A7 A2 A3 A9

T
A8 N1 N1 N2 A4 N2 N2 A10

IP
N5 N3 N3 N4 N4 N6

(a)

CR
1
(5) (6)
8 9

(1)
7
5
US 6
10

(4)
AN
2 (2) 3 (3) 4

(b)
M

Figure 2: Illustration of a genetic programming individual in tree form (a) and in truss
form.
ED

the defined function and terminal sets.

HGP produces the initial population based on two genetic programming


parameters: population size (P S) and initial depth of the trees (di ). It
PT

produces a population of trees with the size of P S with the maximum


length of di ; most common method to generate the initial population
of GP is Ramped-half-and-half method (Koza, 1992).
CE

3. Decoding: In genetic programming, each tree in the initial population


represents a candidate solution and in truss optimization problem, it
represents a truss. To achieve a real truss representation, HGP decodes
AC

each tree into a structure. Soh and yang (Soh and Yang, 2000) proposed
an efficient mapping scheme for this purpose. This scheme preserves
the mandatory feature of decoding a tree to a truss in 1-to-1 mapping
and is fully described in (Soh and Yang, 2000) . HGP decodes each of
the trees into a truss by the provided scheme. Figure 2 (b) depicts a
sample tree decoded into truss format by this mapping method.

11
ACCEPTED MANUSCRIPT

4. Weight calculation: After decoding, genetic programming measures


the fitness of each individual based on the objective function. In this
step, HGP evaluates the structural cost of a truss. It calculates the
weight of obtained trusses (or decoded trees) according to Equation 1.
HGP calculates the structural weight by evaluating the weight of each

T
member and cost of each node but, lumped masses are excluded.

IP
5. Static and modal analysis: After assigning the cost of each truss,
HGP employs static and dynamic analysis on each truss to penalize
each truss if it violates any design considerations. HGP uses OpenSEES

CR
to perform the finite element analysis. OpenSEES provides various
types of elements and solvers to handle complex structural problems(Mazzoni
et al., 2006).

US
6. Genetic operators (Crossover and Mutation): Genetic operators
empower genetic programming to explore the search space and excel the
competence of its population through the evolution process. Crossover
operator takes two parents from the initial population based on a pro-
AN
portionate fitness-based method such as tournament and probability
of crossover (pc ). It cuts each parent from a random location of the
tree and swaps the sliced fragments of two parents with each other to
M

generate two news offsprings. This operator leads to generation of trees


with more variety and may result in to find a more competent solution
for the optimization problem.
ED

Mutation operator takes one parent based on a proportionate fitness-


based method and probability of mutation (pm ). It cuts it from a
random point and plants a subtree in the cut point to produce an off-
spring. It leads HGP to evade the local optimums and traps in complex
PT

optimization problems during the evolution process, especially in the


later generations. If the depth of any produced offspring exceeds the
maximum depth of trees (dm ), genetic programming aborts that tree
CE

and replaces one of its parents with it (Koza, 1992).


7. Nelder-Mead operator: Genetic programming can produce various
trees in the initial population, but the numeric elements remain un-
AC

changed if no mathematical operator is present in the function set.


This pitfall leads to stationary design variables or cross-section areas
in truss optimization problem. Assimi et al. proposed an alteration
operator to handle this drawback (Assimi et al., 2017) but it is also
dependent on a random variable, and it is not highly efficient in very
non-convex problems such as truss optimization with static and dy-

12
ACCEPTED MANUSCRIPT

namic constraints. To address this issue, we proposed a local search


operator based on the Nelder-Mead downhill simplex method.
The Nelder-Mead method such as other local search strategies is in-
dependent of evaluating the derivation of objective functions (Nelder
and Mead, 1965). It employs simplex polytope of N+1 dimensions in

T
an N-dimensional problem. Each vertex of this simplex can be shown

IP
as a vector, and it evaluates the objective function on each of its ver-
tices (test points). Then it sorts the test points in order that N + 1th
test point has the worst value in optimization problem. In next step,

CR
Nelder-Mead finds the centroid of the simplex (x̄). Nelder-Mead em-
ploys the reflection operator and finds a new point (xr ) based on the
centroid and the worst point. Based on some conditions, it employs

US
other operators such as expansion, contraction and reduction to im-
prove the objective function value in the test points. This process goes
on till a termination criterion such as reaching a threshold in the test
points gets satisfied. More information about the Nelder-Mead proce-
AN
dure and its functions can be found in (Nelder and Mead, 1965, Han
and Neumann, 2006, Yang, 2010) Figure 3 depicts a schematic of the
Nelder-Mead operators to improve the solutions.
M

HGP takes a parent based on a proportionate fitness-based method


such as tournament and probability of Nelder-Mead operation (pnm ).
This operator takes one parents and preserves the topology of the truss
ED

but optimizes the cross-section areas (function members of the tree)


to achieve a better solution. It uses Matlab Optimization Toolbox
(Branch and Grace, 1998) for this purpose and produces offsprings
with improved design variables. It is obvious that HGP only employs
PT

this operator for problems with continuous design variables and other
problems with discrete design variables require no change in their cross-
section areas since they choose cross-sections from a set of predefined
CE

available cross-sections.
8. Termination criterion and designation of final result: After per-
forming genetic operators and local search operator on its population,
AC

HGP accumulates all of the offsprings of the operators in a pool. It se-


lects P S trees from the pool and generates the next population. This
next population replaces the initial population, and it iterates till it
reaches the termination criterion. Hence, it terminates and designates
the best-found solution through the evolution process as the solution.

13
ACCEPTED MANUSCRIPT

x2

T
V2

xe

IP
x2
xr
x1

CR
x VN+1
xc xN+1
xN+1
x1

(a)
US (b)

Figure 3: Illustration of Nelder-Mead operators: (a) reflection, expansion and contractions,


(b) reduction
AN
Figure 4 depicts the pseudo-code of the proposed hybrid genetic program-
ming and Nelder-Mead algorithm and its procedures.
M

4. Numerical examples
ED

This section investigates the performance of HGP and applies it to some


truss benchmark examples for topology and sizing optimization. This prob-
lem subjects to static and dynamic constraints such as structural kinematic
stability, maximum allowable stress in elements, critical buckling loading,
PT

maximum allowable displacement in nodes and natural frequency limits. For


all examples, both types of continuous and discrete design variables are con-
sidered, and trusses undergo multiple load conditions. The cost function is
CE

the structural weight of the truss including the weight of the members and
cost of each node; but, it excludes the lumped masses into the calculation.
The members are supposed as tabular with buckling coefficient (ki ) as 4 in
AC

Equation 2 and cost of each node (cj ) is assumed as 5 kg (Kaveh and Zol-
ghadr, 2013). The results obtained by HGP are compared to reports available
in the literature (Xu et al., 2003, Kaveh and Zolghadr, 2013, Savsani et al.,
2016, 2017) .

Genetic programming control parameters are as follows: The population

14
ACCEPTED MANUSCRIPT

Hybrid genetic programming with Nelder-Mead

Set control parameters (P S, pc , pm , pnm , G, di , dm )


Generate random initial population:
Pop = {P1 , P2 , . . . , PP S }
Gen = 1
while Gen ≤ G do

T
for i to P S do
Decode each individual
Evaluate fitness of each individual

IP
Select parents
if rand ≤ pc then
Perform crossover

CR
end if
if rand ≤ pm then
Perform mutation
end if
if rand ≤ pnm then
Perform Nelder-Mead local search operator:

US
while Termination criterion not satisfied do
Generate a simplex polytope with N + 1 vertices in N dimensions
Evaluate objective function at vertices
Reorder the vertices based on its rank
f (x1 ) ≤ f (x2 ) · · · ≤ f (xN +1 )
AN
Measuring the centroid of the simplex
Perform reflection on the worst vertex and finding new point (xr )
if f (x1 ) ≤ f (xr ) < f (xN ) then
xN +1 ← xr
end if
M

if f (xr ) < f (x1 ) then


Perform expansion on (xr ) and finding new point (xe )
if f (xe ) < f (xr ) then
xN +1 ← xe
else
ED

xN +1 ← xr
end if
if f (xr ) > f (xN ) then
Perform contraction to find the new point (xc )
end if
PT

if f (xc ) < f (xN +1 ) then


xN +1 ← xc
else
Perform reduction
end if
CE

end if
end while
end if
Form a pool including operators offspring
Select population for the next generation (NewPop)
AC

Pop ← NewPop
Gen = Gen + 1
end for
end while
Return The best individual

Figure 4: The pseudo-code of hybrid genetic programming with Nelder-Mead

15
ACCEPTED MANUSCRIPT

size (P S) is 200, crossover probability (pc ) as 80 %, mutation probability


(pm ) as 10 %, Nelder-Mead local search operator probability (pnm ) as 30%.
The termination criterion of the algorithm is the maximum number of itera-
tions (G) as 500; initial depth (di ) and maximum depth (dm ) of the trees are
6 and 15, respectively. Because GP is an stochastic evolutionary algorithm,

T
30 times the algorithm has been run independently to obtain robust and re-

IP
liable results; the best run has been presented as the final result of HGP.
All problems are implemented in the Matlab environment and finite element
analysis is performed by OpenSEES. The codes are executed on a computer

CR
having configuration: Core 2 Duo CPU @2.53 GHz, RAM 4 GB. Genetic pro-
gramming is a computationally expensive method; the most expensive parts
include generation of the initial population, evaluation the fitness of popu-

US
lation, performing secondary operators and selection of individuals for the
next generation. Employing a local search operator may increase the total
computational time of the algorithm but empowers it to converge to a better
solution. The obtained results by HGP are also compared with the SOGP al-
AN
gorithm introduced in (Assimi et al., 2017). It should be noticed that in truss
optimization problems subject to static and dynamic constraints with dis-
crete variables, no change in cross-sections is feasible because the optimizer
M

should choose the cross-sections from a set of pre-defined cross-sections. Be-


cause of that, HGP and SOGP do not employ their customized operators:
Nelder-Mead local search operator and Alteration operator, respectively. Ac-
ED

cordingly, the results of both algorithm will be identical and only one has
been reported in the corresponding tables.

4.1. Example 1: 24-bar planar truss


PT

Figure 5 depicts its ground structure. It consists of 24 members and eight


nodes which in 6 nodes are essential. It bears a nonstructural lumped mass
on node 3 with 500 kg. This problem has been studied for two different types
CE

of design variables; continuous and discrete cross-sections classify this prob-


lem into two cases, and each includes different material properties, loading
conditions, cross-section area limits and the maximum allowable stress (Xu
AC

et al., 2003, Savsani et al., 2017). However, both cases are common in some
design considerations: both cases subject to the maximum allowable verti-
cal displacement with 10 mm in nodes 5 and 6 (δ5y&6y ); They also consider
the limit on the first natural frequency with f1 ≥ 30 Hz as the dynamic
constraint.

16
ACCEPTED MANUSCRIPT

2.5 m 2.5 m 2.5 m


(3) F1
12
(2)

1.5 m
13
F2 11 16 17
6 7 9 14
5

T
(1) (4)
4 8 10
2 15
3

IP
18 20
19

2.5 m
1 21

CR
Y
(7) (8)
X 22 (5) 23 (6) 24

Added mass
F2

US
Figure 5: the ground structure of 24-bar planar truss.
F1
AN
4.1.1. Case 1: continuous variables
Truss members subject to the maximum allowable stress of 172.43 MPa
M

(Xu et al., 2003); the modulus of elasticity and density are 6.9 × 1010 Pa
and 2740 kg/m3 , respectively. Cross-section areas ranges from 1 cm2 to 40
cm2 and the truss undergoes two multiple load conditions which are listed in
ED

Table 1.
Figure 6 illustrates the improvement of the best result obtained by HGP
during the evolution process. It initiated with 453.9534 kg and converged
with 119.0855 kg at iteration 78. Figure 7 represents the solution topology
PT

Table 1: Loading conditions for all problems


CE

F1 F2
24-bar Case1 Load condition 1 5 × 104 N 0N
Load condition 2 0 N 5 × 104 N
AC

Case2 Load condition 1 10 × 104 N 0N


Load condition 2 0N 10 × 104 N
20-bar All cases Load condition 1 5 × 105 N 0N
Load condition 2 0N 5 × 105 N

17
ACCEPTED MANUSCRIPT

500

PSO (Kaveh and Zolghadr, 2013)


450
TLBO (Savsani et al., 2016)
CSS (Kaveh and Zolghadr, 2013)
400

T
MS-TLBO (Savsani et al., 2016)
This Study
Best weight (kg)

350

IP
300

CR
250

200

150

100
0 50 100 150
US
200 250 300 350 400 450 500
AN
Iteration

Figure 6: The improvement of the best solution during the run for 24-bar planar truss
case 1.
M

obtained by HGP; it contains nine truss members and six nodes; in other
words, HGP identified 14 truss members and two nodes as redundant ele-
ED

ments in the design space.

Table 2 lists the cross-section areas, structural weight, nodal displace-


PT

ments, and the fundamental frequency of the solution design; it also compares
the obtained result by HGP with other results reported in the literature. It
states that HGP found a lighter truss than 1-D search method (Xu et al.,
CE

2003), PSO and CSS (Kaveh and Zolghadr, 2013), TLBO (Savsani et al.,
2016) and MS-TLBO (Savsani et al., 2016). HGP improved the best-known
result obtained by MS-TLBO by replacing member 9 with a larger cross-
AC

section area and member 13 with a smaller cross-section area.

4.1.2. Case 2: discrete design variables


Truss members subject to the maximum allowable stress of 180 MPa
and the modulus of elasticity and density are 2 × 1011 Pa and 7860 kg/m3 ,

18
ACCEPTED MANUSCRIPT

12

16
7 14
9
8

T
15

IP
CR
22 24

US
Figure 7: Optimal truss for 24-bar planar truss case 1 and case 2.

Table 2: comparison on optimal design parameters for the 24-bar planar truss case 1 by
various algorithms
AN
1-D search CSS PSO TLBO MS-TLBO SOGP HGP
method (Xu et al., 2003) (Kaveh and Zolghadr, 2013) (Kaveh and Zolghadr, 2013) (Savsani et al., 2016) (Savsani et al., 2016) (Assimi et al., 2017) This study
2
A7 (cm ) 11.0000 19.2000 20.1000 19.0214 18.9992 19.0214 18.9992
A8 (cm2 ) 9.5100 3.0000 14.8000 2.9134 2.8997 2.9134 2.8997
A9 (cm2 ) 15.0000 1.4000 Removed 1.7381 1.7727 1.5271 1.7738
A12 (cm2 ) 36.5000 4.0000 2.4000 4.5314 4.5233 4.5014 4.5233
A13 (cm2 ) Removed 14.1300 14.9000 13.8208 13.8848 13.8100 13.8838
M

A14 (cm2 ) 17.6000 Removed 1.2000 Removed Removed Removed Removed


A15 (cm2 ) 13.8000 3.3000 6.5000 2.8950 2.8566 2.8950 2.8566
A16 (cm2 ) 16.0000 23.9000 23.9000 23.8920 23.8830 23.8920 23.8830
A22 (cm2 ) Removed Removed Removed Removed 1.0000 Removed 1.0000
A23 (cm2 ) Removed 1.0400 4.7000 1.1296 Removed 1.2021 Removed
A24 (cm2 ) Removed 1.4000 22.1000 1.3115 1.0786 2.1923 1.0786
Best weight (kg) 167.0 119.75 151.63 119.1304 118.9367 119.5060 118.9361
f1 (Hz) 30.0000 30.0000 30.0000 30.0154 30.0000 30.0022 30.0000
ED

δ5y (mm) 3.2000 8.6000 1.2000 8.9896 9.0023 8.9831 9.0244


δ6y (Hz) 3.0000 8.9000 8.6000 9.8781 9.2413 9.8885 9.9997

respectively (Savsani et al., 2017). Discrete design variables are selected from
PT

a set of available cross-section areas from the American Institute of Steel


Construction (AISC) code which are listed in Table 3. The truss undergoes
two multiple load conditions which are listed in Table 1.
CE

Figure 8 illustrates the improvement of the best result obtained by HGP


during the evolution process. It started with 872.6595 kg and converged
with 225.2700 kg at iteration 328. Figure 7 represents the solution topology
AC

obtained by HGP which is identical to the solution layout found in the case 1.

Table 4 lists the design parameters of the solution design; it also compares
the obtained result by HGP with other results reported in the literature. It
states that HGP found a lighter truss than all various algorithms presented
in (Savsani et al., 2017) without violating the constraints and the design

19
ACCEPTED MANUSCRIPT

Table 3: Available cross-sectional areas from the AISC design code.


No. cm2 No. cm2
1 0.7161 33 24.7741

T
2 0.9097 34 24.9677

IP
3 1.2645 35 25.0322
4 1.6129 36 26.9677
5 1.9806 37 27.2258

CR
6 2.5226 38 28.9677
7 2.8516 39 29.6128
8 3.6323 40 30.9677
9
10
11
3.8839
4.9419
5.0645
US41
42
43
32.0645
33.0322
37.0322
AN
12 6.4129 44 46.5806
13 6.4516 45 51.4193
14 7.9226 46 55.0322
15 8.1677 47 59.9999
M

16 9.4000 48 69.9999
17 10.0839 49 74.1943
18 10.4516 50 87.0966
ED

19 11.6129 51 89.6772
20 12.8387 52 91.6127
21 13.7419 53 99.9998
22 15.3548 54 103.2256
PT

23 16.9032 55 109.0320
24 16.9677 56 121.2901
25 18.5806 57 128.3868
CE

26 18.9032 58 141.9352
27 19.9354 59 147.7416
28 20.1935 60 158.0642
AC

29 21.8064 61 170.9674
30 22.3871 62 180.6448
31 22.9032 63 193.5480
32 23.4193 64 216.1286

20
ACCEPTED MANUSCRIPT

1000

920
This Study
840

760

T
Best weight (kg)

680

IP
600

520

CR
440

360

280

200
0 50 100 150

US
200 250
Iteration
300 350

Figure 8: The improvement of the best solution during the run for 24-bar planar truss
400 450 500
AN
case 2.

considerations. The obtained solution topology is identical to the topology


M

obtained by MHTS, PVS, and MPVS algorithms; but, HGP selected differ-
ent cross-section areas for members 8, 9, 12, 22 and 24 and provided a better
solution. Table 4 also shows that HGP demonstrated reasonable computa-
ED

tional time with population size of 200 in comparison with other methods in
(Savsani et al., 2017) with population size of 50.
PT
CE
AC

21
AC
CE
Table 4: comparison on optimal design parameters for the 24-bar planar truss case 2 by various algorithms
TLBO MTLBO MHTS MWWO PVS MPVS HGP
(Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) This study
PT
A2 (cm2 ) 0.7161 Removed Removed Removed Removed Removed Removed
A5 (cm2 ) 0.7161 Removed Removed Removed Removed Removed Removed
2
A7 (cm ) 18.5806 18.5806 18.5806 18.5806 18.9032 18.5806 18.5806
A8 (cm2 ) 6.4516 6.4129 6.4516 6.4129 6.4129 6.4516 6.4129
ED
A9 (cm2 ) 6.4129 3.8839 1.9806 6.4129 1.2645 1.6129 1.2645
A12 (cm2 ) 3.8839 2.5226 2.5226 3.6323 2.5226 2.5226 2.8516
A14 (cm2 ) Removed 1.2645 2.5226 Removed 2.8516 2.5226 2.5226
2
A15 (cm ) 3.6323 4.9419 6.4129 3.6323 6.4129 6.4129 6.4129
M

22
A16 (cm2 ) Removed Removed 19.9354 Removed 19.9354 19.9354 19.9354
A17 (cm2 ) 10.0839 10.0839 Removed 10.0839 Removed Removed Removed
A19 (cm2 ) 0.7161 Removed Removed Removed Removed Removed Removed
A20 (cm2 ) 6.4129 6.4516 Removed 6.4129 Removed Removed Removed
2
A21 (cm ) 10.4516 10.0839 Removed 10.0839 Removed Removed Removed
A22 (cm2 ) Removed 0.9097 1.6129 0.7161 1.6129 1.6129 1.2645
AN
A24 (cm2 ) 8.1677 7.9226 0.9097 8.1677 0.9097 1.2645 1.6129
Best weight (kg) 243.2922 227.3884 226.3047 229.0035 226.1573 225.8168 225.2700
f1 (Hz) 30.1568 30.0719 30.0763 30.2718 30.0048 30.0152 30.3033
δ5y (mm) 3.2037 2.8467 2.3577 3.2142 2.1874 2.3277 2.2739
ACCEPTED MANUSCRIPT

δ6y (Hz) 3.7693 3.4258 3.3273 3.7777 3.3623 3.3282 3.3513


US
Mean time (sec) 25.14 19.27 25.38 16.71 23.24 19.33 20.851
CR
IP
T
ACCEPTED MANUSCRIPT

4.2. Example 2: 20-bar planar truss


Figure 9 depicts its ground structure. It consists of 20 members and
nine nodes. Two nodes (node 1 and 9) support the truss, node 4 bears the
loadings and no lumped mass added on the truss nodes. It has been studied
for two different types of design variables and two different displacement

T
limitations. Case 1 and 2 discuss this benchmark by employing continuous

IP
design variables but with different maximum allowable displacement in nodes.
Case 3 considers the discrete design variables to be selected from a set of
available cross-section areas. All cases are subject to a common dynamic

CR
constraint for the first and second natural frequencies as follows: f1 ≥ 60
and f2 ≥ 100 (Xu et al., 2003, Savsani et al., 2017). Xu et al. (2003) applied
their algorithm to this benchmark with nodal displacement constraint of 10

US
mm in node four. Inspection of its solution design shows that it violates the
nodal displacement constraint with 58 mm. Kaveh et al. reconfigured this
benchmark into two cases, one with the nodal displacement of 10 mm and
AN
the other with the deflection of 58 mm (Kaveh and Zolghadr, 2013).

4.2.1. Case 1: continuous variables with nodal displacement of 10 mm


Truss members subject to the maximum allowable stress of 172.43 MPa
M

(Xu et al., 2003) and the modulus of elasticity and density are 6.9 × 1010 Pa

2m 2m
ED

(1) 1 (2) 2 (3)

13 15
2m
PT

7 8 9

14 16
(4)
3 4
(6) F1
CE

(5)
17 19 F2
2m

10 11 12
AC

18 20

5 6
(9) (8) (7)
Added mass

Figure 9: the ground structure of 20-bar planar truss.

23
ACCEPTED MANUSCRIPT

750

PSO (Kaveh and Zolghadr, 2013)


700
TLBO (Savsani et al., 2016)
650 CSS (Kaveh and Zolghadr, 2013)
MS-TLBO (Savsani et al., 2016)

T
600 This Study
Best weight (kg)

IP
550

500

CR
450

400

350

300
0 50 100 150
US
200 250 300 350 400 450 500
AN
Iteration

Figure 10: The improvement of the best solution during the run for 20-bar planar truss
case 1.
M

and 2740 kg/m3 , respectively. Cross-section areas ranges from 1 cm2 to 100
cm2 and the truss undergoes two multiple load conditions which are listed in
ED

Table 1. The nodal displacement constraint for this case is 10 mm on node


4 in the y-direction.
Figure 10 illustrates the improvement of the best result obtained by HGP
PT

during the evolution process. It initiated with 650.4960 kg and converged


with 315.9151 kg at iteration 62. Figure 11 represents the solution topology
obtained by HGP; it consists of 8 truss members and six nodes; in other
CE

words, HGP identified 12 truss members and three nodes as redundant ele-
ments in the design space.

Table 5 lists the design parameters of the solution design; it also compares
AC

the obtained result by HGP with other results reported in the literature.
It states that HGP found a better result than PSO and CSS (Kaveh and
Zolghadr, 2013), TLBO and MS-TLBO (Savsani et al., 2016) and provided
a lighter truss.

24
ACCEPTED MANUSCRIPT

13 15

T
IP
CR
11

18

5
US 20
AN

Figure 11: Optimal truss for 20-bar planar truss case 1 and case 2.
M

Table 5: comparison on optimal design parameters for the 20-bar planar truss case 1 by
various algorithms
CSS PSO TLBO MS-TLBO SOGP HGP
ED

(Kaveh and Zolghadr, 2013) (Kaveh and Zolghadr, 2013) (Savsani et al., 2016) (Savsani et al., 2016) (Assimi et al., 2017) This study
2
A1 (cm ) 39.9200 44.0500 44.6461 43.3431 45.2592 43.4298
A5 (cm2 ) 49.2800 42.0800 42.3940 43.3459 42.4262 43.3459
A8 (cm2 ) 46.4100 42.1800 43.0649 43.4298 43.3431 43.3431
A11 (cm2 ) 41.2000 46.7500 45.5652 43.3709 46.2942 43.3431
A13 (cm2 ) 59.0300 63.6100 59.1147 61.3545 60.2075 61.6438
A15 (cm2 ) 63.1500 59.5400 66.8233 61.9340 61.3404 61.6438
PT

A18 (cm2 ) 64.9700 53.4600 61.6249 61.6438 61.2757 61.6438


A20 (cm2 ) 58.4000 71.5100 57.9915 61.3404 61.3404 61.3404
Best weight (kg) 317.1900 318.2300 316.5692 315.9310 316.3973 315.9151
f1 (Hz) 120.0000 115.1000 118.2541 119.4168 119.1374 119.5810
f2 (Hz) 192.1000 186.9000 190.6130 190.9873 192.2143 191.1522
δ4y (mm) 10.0000 10.0000 9.9999 9.9995 9.9890 10.0000
CE

4.2.2. Case 2: continuous variables with nodal displacement of 58 mm


All design considerations are identical to case 1, but nodal displacement
AC

is different. This problem subject to a nodal displacement of 58 mm in node


four along the y-axis. Figure 12 illustrates the improvement of the best result
obtained by HGP during the evolution process. It initiated with 543.4184 kg
and converged with 150.6381 kg at iteration 82. The obtained topology is
identical to what HGP found in case 1, but it is notable that the cross-section

25
ACCEPTED MANUSCRIPT

740

680 PSO (Kaveh and Zolghadr, 2013)


TLBO (Savsani et al., 2016)
620
CSS (Kaveh and Zolghadr, 2013)
560 MS-TLBO (Savsani et al., 2016)

T
This Study
Best weight (kg)

500

IP
440

CR
380

320

260

200

140
0 50 100 150
US
200 250 300 350 400 450 500
AN
Iteration

Figure 12: The improvement of the best solution during the run for 20-bar planar truss
case 2.
M

Table 6: comparison on optimal design parameters for the 20-bar planar truss case 2 by
various algorithms
1-D search CSS PSO TLBO MS-TLBO SOGP HGP
ED

method (Xu et al., 2003) (Kaveh and Zolghadr, 2013) (Kaveh and Zolghadr, 2013) (Savsani et al., 2016) (Savsani et al., 2016) (Assimi et al., 2017) This study
2
A1 (cm ) 53.3800 14.7400 15.6700 15.1964 14.4998 16.6133 14.4987
A4 (cm2 ) Removed Removed 12.7600 Removed Removed Removed Removed
A5 (cm2 ) 3.0200 19.2200 18.4100 19.2765 19.0352 19.0352 19.0347
A8 (cm2 ) 53.3800 19.3300 22.3400 19.2599 19.0419 19.0352 19.0347
A11 (cm2 ) 3.0400 19.0300 14.6600 19.3129 19.0366 19.0352 19.0347
A13 (cm2 ) 42.1200 20.5900 19.1900 20.9239 20.5043 20.5043 20.5042
A15 (cm2 ) 71.9700 20.5500 26.6600 20.5347 20.5048 20.5043 20.5042
PT

A18 (cm2 ) 53.4400 32.4000 33.8500 32.1239 32.0125 32.0125 32.0125


A20 (cm2 ) 3.7600 32.6400 31.1700 33.0428 32.0139 32.0139 32.0125
Best weight (kg) 225.9000 151.9200 161.8800 152.6625 150.6455 151.7989 150.6381
f1 (Hz) 60.0000 116.6000 108.6000 116.6909 116.8936 116.8542 116.8950
f2 (Hz) 130.4000 185.1000 198.6000 185.9479 185.7140 188.7231 185.7110
δ4y (mm) 58.0000 24.0000 24.0000 24.2009 24.6063 24.2893 24.6077
CE

areas are different.

Table 6 lists the design parameters of the solution design and asserts that
AC

HGP found a lighter truss in comparison with other methods in the literature
such as PSO and CSS (Kaveh and Zolghadr, 2013), TLBO and MS-TLBO
(Savsani et al., 2016).

26
ACCEPTED MANUSCRIPT

Table 7: comparison on optimal design parameters for the 20-bar planar truss case 3 by
various algorithms
TLBO MTLBO MHTS MWWO PVS MPVS HGP
(Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) This study
2
A1 (cm ) 13.7419 13.7419 15.3548 20.1935 16.9032 15.3548 15.3548
A2 (cm2 ) 0.7161 Removed Removed Removed Removed Removed Removed
A4 (cm2 ) 13.7419 13.7419 13.7419 16.9032 13.7419 13.7419 13.7419

T
A5 (cm2 ) 16.9677 16.9032 16.9032 16.9032 13.7419 15.3548 15.3548
A8 (cm2 ) 15.3548 13.7419 13.7419 22.3871 16.9032 15.3548 15.3548
A9 (cm2 ) 0.7161 Removed Removed Removed Removed Removed Removed
A11 (cm2 ) 16.9032 16.9032 16.9677 12.8387 13.7419 15.3548 15.3548

IP
2
A13 (cm ) 22.9032 21.8064 22.3871 18.5806 19.9354 21.8064 21.8064
A15 (cm2 ) 18.5806 19.9354 18.5806 21.8064 21.8064 19.9354 19.9354
A18 (cm2 ) 19.9354 20.1935 18.9032 22.3871 22.3871 19.9354 19.9354
A20 (cm2 ) 21.8064 22.3871 23.4193 18.9032 20.1935 22.3871 22.3871
Best weight (kg) 342.8615 335.4113 335.7534 351.8429 335.4113 335.0404 335.0404

CR
f1 (Hz) 64.9482 65.3759 65.2803 65.6465 65.4716 65.3335 65.3335
f2 (Hz) 114.3366 114.3824 114.3286 116.9238 114.4201 114.4593 114.4593
δ4y (mm) 9.9991 9.9909 9.9986 9.7684 9.9909 9.9951 9.9951
Mean time (sec) 28.76 22.56 26.91 11.76 27.97 22.51 29.92

4.2.3. Case 3: discrete design variables


Truss members subject to the maximum allowable stress of 180 MPa
and the modulus of elasticity and density are 2 × 1011 Pa and 7860 kg/m3 ,
US
AN
respectively (Savsani et al., 2017). Discrete design variables are selected from
Table 3. The truss undergoes two multiple load conditions identical to load
conditions in cases 1 and 2.
Figure 13 illustrates the improvement of the best result obtained by HGP
M

during the evolution process. It initiated with 1532.3240 kg and converged


with 335.0404 kg at iteration 240. Figure 14 represents the solution topology
obtained by HGP; it contains nine truss members and six nodes; in other
ED

words, HGP identified 11 truss members and two nodes as redundant ele-
ments in the design space.

Table 7 lists the design parameters of the solution design; it compares


PT

the solution design found by HGP with other results reported in the litera-
ture. It states that HGP reached the solution identical to what reported by
MPVS (Savsani et al., 2017) in topology and sizing of cross-section areas;
CE

it seems that it is the global optimum for this problem; it also provided a
lighter truss in comparison with other algorithms excluding MPVS. Table
7 also shows that HGP demonstrated reasonable computational time with
AC

population size of 200 in comparison with other methods in (Savsani et al.,


2017) with population size of 50.

4.3. Example 3: 72-bar spatial truss


Figure 15 depicts its ground structure. It consists of 72 members and 20
nodes which in 8 nodes are essential. These 72 members classified into 16

27
ACCEPTED MANUSCRIPT

900

850
This Study
800

750

T
700
Best weight (kg)

IP
650

600

550

CR
500

450

400

350

300
1 50 100 150
US
200 250 300 350 400 450 500
AN
Iteration

Figure 13: The improvement of the best solution during the run for 20-bar planar truss
case 3.
M

groups using symmetry. The nonstructural lumped mass on top nodes (nodes
1-4) is 2270 kg. This problem has been studied for two different types of
ED

design variables. Continuous and discrete cross-sections classify this problem


into two cases, and each includes different material properties, cross-section
areas range and the maximum allowable stress (Xu et al., 2003, Savsani
PT

et al., 2017). However, both cases subject to maximum allowable vertical


displacement with 6.35 mm in top nodes along both x and y directions;
They also impose limits on the first and third natural frequency constraints
CE

as follows: f1 ≥ 4 Hz and f3 ≥ 6 Hz (Kaveh and Zolghadr, 2013, Savsani


et al., 2017)

4.3.1. Case 1: continuous variables


AC

Truss members subject to the maximum allowable stress of 172.375 MPa


(Kaveh and Zolghadr, 2013) and the modulus of elasticity and density are
6.895 × 1010 Pa and 2767.99 kg/m3 , respectively. Cross-section areas ranges
from 1 cm2 to 30 cm2 and the truss undergoes two multiple load conditions
which are listed in Table 8.

28
ACCEPTED MANUSCRIPT

T
13 15

IP
8

CR
4

US 11
AN
18 20

5
M

Figure 14: Optimal truss for 20-bar planar truss case 3.


ED
PT

Table 8: The loading condition for 72-bar spatial truss


Load condition 1 Load condition 2
CE

Nodes Px (kN) Py (kN) Pz (kN) Px (kN) Py (kN) Pz (kN)


1 22.25 22.25 -22.25 0.0 0.0 -22.25
AC

2 0.0 0.0 0.0 0.0 0.0 -22.25


3 0.0 0.0 0.0 0.0 0.0 -22.25
4 0.0 0.0 0.0 0.0 0.0 -22.25

29
ACCEPTED MANUSCRIPT

T
IP
(4)
16
18 15

CR
11 4
(1) 17 9
13 (8) 10 (3)
1.524 m

12
1 14
3

(5)
6
US
(2)
(12)
7
8
(7)
1.524 m

5 2
19
AN
21

(9) (6)
(16) (11)
1.524 m

20
M

37
39

(13) (10)
ED

(20) (15)
1.524 m

38
55
57

(17)
PT

(14)
(19)
56
3.0
48 8m Z
m 3.04
Y
CE

Added mass (18)


X

Figure 15: the ground structure of 72-bar spatial truss.


AC

30
ACCEPTED MANUSCRIPT

1000
950 PSO (Kaveh and Zolghadr, 2013)
900 TLBO (Savsani et al., 2016)
CSS (Kaveh and Zolghadr, 2013)
850
MS-TLBO (Savsani et al., 2016)
800

T
This Study
Best weight (kg)

750

IP
700
650

CR
600

550

500
450

400
0 50 100 150
US
200 250
Iteration
300 350 400 450 500
AN
Figure 16: The improvement of the best solution during the run for 72-bar spatial truss
case 1

Figure 16 illustrates the improvement of the best result obtained by HGP


M

during the evolution process. It started with 1401.1452 kg and converged


with 435.0876 kg at iteration 68. Figure 17 represents the solution topology
obtained by HGP; it contains 54 truss members categorized into ten groups
ED

and 20 nodes; in other words, HGP identified 18 truss members or 6 groups


as redundant elements in the design space.
Table 9 lists the design parameters of the solution design; it also compares
the obtained result by HGP with other results reported in the literature.
PT

It states that HGP found a lighter truss than PSO and CSS (Kaveh and
Zolghadr, 2013), TLBO and MS-TLBO (Savsani et al., 2016). The evident
difference between the deflection of node four along x-direction for the so-
CE

lution design obtained by HGP and MS-TLBO is because of that Tejani et


al. reported the deflection only based on the first loading condition, but
actually, this node deflects more under the second load condition. Figure 18
AC

depicts the displacement in nodes of the solution design, and it states that
no violation of nodal displacement occurred.

31
ACCEPTED MANUSCRIPT

(1) (2)

T
(5) (6)

IP
CR
(9) (10)

US (13) (14)

Z
AN
X
(17) (18)

Figure 17: Optimal truss for 72-bar spatial truss case 1.


M
ED

Table 9: comparison on optimal design parameters for the 72-bar spatial truss case 1 by
various algorithms
Group Members CSS PSO TLBO MS-TLBO SOGP HGP
(Kaveh and Zolghadr, 2013) (Kaveh and Zolghadr, 2013) (Savsani et al., 2016) (Savsani et al., 2016) (Assimi et al., 2017) This study
PT

G1 (cm2 ) A1 − A4 5.5400 5.3000 4.9192 5.1683 4.9953 5.0018


G2 (cm2 ) A5 − A12 8.0600 6.9800 10.2376 10.2401 10.2401 10.2377
G3 (cm2 ) A13 − A16 Removed 5.6000 Removed Removed Removed Removed
G4 (cm2 ) A17 − A18 9.0400 13.5600 Removed Removed Removed Removed
G5 (cm2 ) A19 − A22 8.0700 5.1600 10.0541 9.1723 11.9667 9.4594
G6 (cm2 ) A23 − A30 8.0400 9.4800 7.2706 7.2850 7.2850 7.2925
G7 (cm2 ) A31 − A34 3.1300 Removed Removed Removed Removed Removed
CE

G8 (cm2 ) A35 − A36 Removed Removed 3.7589 3.8942 3.8942 3.8892


G9 (cm2 ) A37 − A40 10.0100 22.5800 12.6451 11.9667 11.9667 11.9596
G10 (cm2 ) A41 − A48 8.1500 6.9800 7.6935 7.7058 7.7058 7.6897
G12 (cm2 ) A53 − A54 Removed 5.1100 Removed Removed Removed Removed
G13 (cm2 ) A55 − A58 20.3200 21.1700 15.1415 16.1732 14.8465 16.0580
G14 (cm2 ) A59 − A66 7.9600 9.5600 7.5945 7.5644 7.6114 7.5718
G15 (cm2 ) A67 − A70 Removed Removed 1.8944 1.9819 1.9819 1.9791
Best weight (kg) 437.8500 504.0600 435.4325 435.1404 437.6783 435.0876
AC

f1 (Hz) 4.0000 4.0000 4.0013 4.0001 4.0000 4.0000


f3 (Hz) 6.0000 6.0000 6.0001 6.0001 6.0000 6.0000
δ1x (mm) 4.3000 2.9000 6.1779 6.1469 6.1602 6.1487
δ2x (mm) 3.5000 2.3000 1.6532 1.6498 1.6824 1.6514
δ3x (mm) 1.7000 2.5000 2.5554 2.5931 2.6647 2.6006
δ4x (mm) 1.6000 2.3000 0.8048 0.8079 1.4000 1.4255
δ1y (mm) 4.3000 2.9000 6.1779 6.1469 6.1602 6.1487
δ2y (mm) 1.6000 2.3000 0.8048 0.8079 0.8810 0.8146
δ3y (mm) 1.7000 2.5000 2.5554 2.5931 2.6647 2.6006
δ4y (mm) 3.5000 2.3000 1.6532 1.6498 1.6824 1.6514

32
ACCEPTED MANUSCRIPT

T
IP
CR
Maximum displacement in x direction for both loadings
Maximum displacement in y direction for both loadings
7

5
US Maximum displacement
6.35 mm
AN
Displacement (mm)

4
M

2
ED

0
PT

1 2 3 4
Node number

Figure 18: Nodal displacements for optimal truss of 72-bar spatial truss case 1.
CE
AC

33
ACCEPTED MANUSCRIPT

4.3.2. Case 2: discrete variables


Truss members subject to the maximum allowable stress of 180 MPa
(Kaveh and Zolghadr, 2013) and the the modulus of elasticity and density
are 2×1011 Pa and 7860 kg/m3 , respectively. Cross-section areas are selected
from Table 3 and the truss undergoes same loading conditions as case 1.

T
Figure 19 illustrates the improvement of the best result obtained by HGP

IP
during the evolution process. It initiated with 1051.6626 kg and converged
with 541.6028 kg at iteration 282. Figure 20 represents the solution topology
obtained by HGP; it consists of 54 truss members categorized into ten groups

CR
and 20 nodes; in other words, HGP identified 18 truss members or 6 groups
as redundant elements in the design space.

US
Table 10 lists the design parameters of the solution design; it also com-
pares the obtained result by HGP with other results reported in the litera-
ture. It states that HGP found a better truss than all other algorithms. It
AN
states that HGP also improved the best-known result obtained by MPVS by
replacing member 1 with a smaller cross-section area and member 9 with a
larger cross-section area.
Figure 21 depicts the displacement in nodes of the solution design, and it
M

states that no violation of nodal displacement occurred. Table 10 also shows


that HGP converged faster than other methods and it is also noticeable that
this experiment conducted with the population size of 200 in comparison
ED

with other methods in (Savsani et al., 2017) with population size of 100.
PT
CE
AC

34
AC
CE
PT
Table 10: comparison on optimal design parameters for the 72-bar spatial truss case 2 by various algorithms
Group Members TLBO MTLBO MHTS MWWO PVS MPVS HGP
(Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) (Savsani et al., 2017) This study
G1 (cm2 ) 2.8516 2.8516 3.6323 2.5226 2.8516 2.8516 2.5226
ED
A1 − A4
G2 (cm2 ) A5 − A12 3.6323 3.6323 3.6323 3.6323 3.6323 3.6323 3.6323
G3 (cm2 ) A13 − A16 2.8516 2.8516 2.8516 2.8516 2.8516 2.8516 2.8516
G4 (cm2 ) A17 − A18 5.0645 5.0645 5.0645 5.0645 5.0645 5.0645 5.0645
G5 (cm2 ) 2.8516 2.8516 2.8516 3.6323 2.8516 2.8516 2.8516
G6 (cm2 )
A19 − A22
3.6323 3.6323
M 3.6323 3.6323 3.6323 3.6323 3.6323

35
A23 − A30
G9 (cm2 ) A37 − A40 4.9419 4.9419 4.9419 4.9419 4.9419 4.9419 5.0645
G10 (cm2 ) A41 − A48 3.6323 3.6323 3.6323 3.6323 3.6323 3.6323 3.6323
G13 (cm2 ) A55 − A58 3.8839 3.8839 3.8839 4.9419 3.8839 3.8839 3.8839
G14 (cm2 ) A59 − A66 3.6323 3.6323 3.6323 3.6323 3.6323 3.6323 3.6323
Best weight (kg) 542.5877 542.5877 546.3281 549.8213 542.5877 542.5877 541.6028
f1 (Hz) 4.0197 4.0197 4.0297 4.2390 4.0197 4.0197 4.0217
AN
f3 (Hz) 6.8520 6.8520 6.8509 6.8518 6.8520 6.8520 6.8525
Max Displacement (mm) 2.9086 2.9086 2.9173 2.6914 2.9086 2.9086 2.8948
Mean time (sec) 99.59 50.86 74.09 33.73 91.58 55.44 37.37
ACCEPTED MANUSCRIPT

US
CR
IP
T
ACCEPTED MANUSCRIPT

1000

950
This Study
900

T
850
Best weight (kg)

800

IP
750

700

CR
650

600

US
550

500
1 50 100 150 200 250 300 350 400 450 500
Iteration
AN
Figure 19: The improvement of the best solution during the run for 72-bar spatial truss
case 2
M

(1) (2)
ED

(5) (6)
PT

(9) (10)
CE

(13) (14)
AC

(18) X
(17)

Figure 20: Optimal truss for 72-bar spatial truss case 2.

36
ACCEPTED MANUSCRIPT

T
IP
CR
Maximum displacement in x direction for both loadings
Maximum displacement in y direction for both loadings
7

6 US Maximum displacement
6.35 mm
AN
5
Displacement (mm)

4
M

2
ED

0
1 2 3 4
PT

Node number

Figure 21: Nodal displacements for optimal truss of 72-bar spatial truss case 2.
CE
AC

37
ACCEPTED MANUSCRIPT

5. Conclusion
In this paper, a hybrid genetic programming (HGP) algorithm proposed
to deal with the topology and sizing optimization problem of trusses sub-
ject to both static and dynamic constraints. HGP produced a random-blind

T
initial population composed of trees; each tree represented a truss. HGP
performed finite element analysis on its population and got its individuals

IP
penalized if each violated any of the design considerations. HGP applied
crossover and mutation to its population to explore the search space and

CR
improve the competence of its population by breeding new offsprings. HGP
also employed a Nelder-Mead local search operator to enhance the true con-
vergence of the algorithm. This operator empowered HGP to look for a
better set of design variables in the vicinity of the found solution. It also

US
provided the benefit of improving all design variables in an iterative process.
The evolution process iterated till a termination criterion got satisfied and
the lightest truss found in evolution designated as the final solution design.
AN
We applied our approach to some numerical examples for both types of con-
tinuous and discrete design variables. Obtained results by HGP compared
with other studies reported in the literature and HGP reached better re-
sults in the most of the cases. The genetic programming natural feature of
M

resembling its individuals in tree form lead to tuning the tree parameters
and structure simultaneously and empowered it to find redundant elements
in the design space. HGP also reached better solutions than SOGP in the
ED

very non-convex problem of truss optimization subject to static and dynamic


constraints and explored the search space more efficiently to obtain a lighter
truss. Some further studies are suggested to develop this approach for non-
PT

deterministic design strategies to incorporate uncertainty in the design. This


study also can be extended to other types of structures such as shells, frames
and composite plates.
CE

Acknowledgment
The authors gratefully acknowledge Prof. Nader Nariman-zadeh for the
AC

stimulating and fruitful discussions.

38
ACCEPTED MANUSCRIPT

References
Assimi, H., Jamali, A., and Nariman-zadeh, N. (2017). Sizing and topology
optimization of truss structures using genetic programming. Swarm and
Evolutionary Computation.

T
Bellagamba, L. and Yang, T. (1981). Minimum-mass truss structures

IP
with constraints on fundamental natural frequency. AIAA Journal,
19(11):1452–1458.

CR
Branch, M. A. and Grace, A. (1998). Optimization Toolbox: for Use with
MATLAB: User’s Guide: Version 1. Math works.

Cheng, M. (2016). A Hybrid Harmony Search algorithm for discrete sizing

US
optimization of truss structure. Automation in Construction, 69:21–33.

Deb, K. (2001). Design of truss-structures for minimum weight using genetic


AN
algorithms. Finite Elements in Analysis and Design, 37(5):447–465.

Deb, K. and Kalyanmoy, D. (2001). Multi-Objective Optimization Using


Evolutionary Algorithms. John Wiley & Sons, Inc., New York, NY, USA.
M

Dorn, W. S. (1964). Automatic design of optimal structures. Journal de


mecanique, 3:25–52.
ED

Fajfar, I., Puhan, J., and Bűrmen, Á. (2017). Evolving a nelder–mead algo-
rithm for optimization with genetic programming. Evolutionary computa-
tion.
PT

Farshchin, M., Camp, C. V., and Maniat, M. (2016a). Multi-class teaching-


learning-based optimization for truss design with frequency constraints.
Engineering Structures, 106(C):355–369.
CE

Farshchin, M., Camp, C. V., and Maniat, M. (2016b). Optimal design of truss
structures for size and shape with frequency constraints using a collabora-
AC

tive optimization strategy. Expert Systems with Applications, 66:203–218.

Fenton, M., McNally, C., Byrne, J., Hemberg, E., McDermott, J., and
O’Neill, M. (2014). Automatic innovative truss design using grammati-
cal evolution. Automation in Construction, 39(C):59–69.

39
ACCEPTED MANUSCRIPT

Fenton, M., McNally, C., Byrne, J., Hemberg, E., McDermott, J., and ONeill,
M. (2016). Discrete planar truss optimization by node position variation
using grammatical evolution. IEEE Transactions on Evolutionary Com-
putation, 20(4):577–589.

T
Gholaminezhad, I., Assimi, H., Jamali, A., and Vajari, D. A. (2016). Uncer-
tainty quantification and robust modeling of selective laser melting process

IP
using stochastic multi-objective approach. The International Journal of
Advanced Manufacturing Technology, 86(5-8):1425–1441.

CR
Gomes, H. M. (2011). Truss optimization with dynamic constraints using a
particle swarm algorithm. Expert Systems with Applications, 38(1):957–
968.

US
Grandhi, R. (1993). Structural optimization with frequency constraints - A
review. AIAA Journal, 31(12):2296–2303.
AN
Grandhi, R. V. and Venkayya, V. B. (1988). Structural optimization with
frequency constraints. AIAA Journal, 26(7):858–866.

Han, L. and Neumann, M. (2006). Effect of dimensionality on the nelder–


M

mead simplex method. Optimization Methods and Software, 21(1):1–16.

Hare, W., Nutini, J., and Tesfamariam, S. (2013). A survey of non-gradient


ED

optimization methods in structural engineering. Advances in Engineering


Software, 59:19–28.

Ho-Huu, V., Nguyen-Thoi, T., Vo-Duy, T., and Nguyen-Trang, T. (2016). An


PT

adaptive elitist differential evolution for optimization of truss structures


with discrete design variables. Computers & Structures, 165(C):59–75.
CE

Jamali, A., Khaleghi, E., Gholaminezhad, I., and Nariman-Zadeh, N. (2016).


Modelling and prediction of complex non-linear processes by using pareto
multi-objective genetic programming. International Journal of Systems
Science, 47(7):1675–1688.
AC

Jamali, A., Khaleghi, E., Gholaminezhad, I., Nariman-zadeh, N., Gho-


laminia, B., and Jamal-Omidi, A. (2014). Multi-objective genetic pro-
gramming approach for robust modeling of complex manufacturing pro-
cesses having probabilistic uncertainty in experimental data. Journal of
Intelligent Manufacturing, 28(1):149–163.

40
ACCEPTED MANUSCRIPT

Kaveh, A. and Ilchi Ghazaan, M. (2015). Optimal design of dome truss struc-
tures with dynamic frequency constraints. Structural and Multidisciplinary
Optimization, 53(3):605–621.

Kaveh, A. and Mahdavi, V. R. (2014). Colliding Bodies Optimization method

T
for optimum discrete design of truss structures. Computers & Structures,
139(C):43–53.

IP
Kaveh, A. and Mahdavi, V. R. (2015). A hybrid CBO-PSO algorithm for

CR
optimal design of truss structures with dynamic constraints. Applied Soft
Computing, 34:260–273.

Kaveh, A. and Zolghadr, A. (2013). Topology optimization of trusses con-

puting, 13(5):2727–2734. US
sidering static and dynamic constraints using the CSS. Applied Soft Com-

Kelner, V., Capitanescu, F., Léonard, O., and Wehenkel, L. (2008). A hybrid
AN
optimization technique coupling an evolutionary and a local search algo-
rithm. Journal of Computational and Applied Mathematics, 215(2):448–
456.
M

Koohestani, B. and Poli, R. (2014). Evolving an improved algorithm for


envelope reduction using a hyper-heuristic approach. IEEE Transactions
on Evolutionary Computation, 18(4):543–558.
ED

Koza, J. R. (1992). Genetic Programming–On the Programming of Computer


Programs by Natural Selection. MIT Press.
PT

Li, L. J., Huang, Z. B., and Liu, F. (2009). A heuristic particle swarm op-
timization method for truss structures with discrete variables. Computers
& Structures, 87(7-8):435–443.
CE

Li, L. J., Huang, Z. B., Liu, F., and Wu, Q. H. (2007). A heuristic particle
swarm optimizer for optimization of pin connected structures. Computers
& Structures, 85(7-8):340–349.
AC

Mazzoni, S., McKenna, F., Scott, M. H., Fenves, G. L., and others (2006).
OpenSees command language manual. Pacific Earthquake Engineering
Research (PEER) Center.

41
ACCEPTED MANUSCRIPT

Miguel, L. F. F. and Miguel, L. F. F. (2012). Shape and size optimization


of truss structures considering dynamic constraints through modern meta-
heuristic algorithms. Expert Systems with Applications, 39(10):9458–9467.

Mortazavi, A. and Toğan, V. (2017). Sizing and layout design of truss struc-

T
tures under dynamic and static constraints with an integrated particle
swarm optimization algorithm. Applied Soft Computing, 51:239–252.

IP
Nelder, J. A. and Mead, R. (1965). A simplex method for function minimiza-

CR
tion. The computer journal, 7(4):308–313.

Poli, R. and Koza, J. (2014). Genetic Programming. In Search Methodologies,


pages 143–185. Springer US, Boston, MA.

US
Savsani, V. J., Tejani, G. G., and Patel, V. K. (2016). Truss topology op-
timization with static and dynamic constraints using modified subpop-
ulation teaching-learning-based optimization. Engineering Optimization,
AN
48(11):1990–2006.

Savsani, V. J., Tejani, G. G., Patel, V. K., and Savsani, P. (2017). Modified
meta-heuristics using random mutation for truss topology optimization
M

with static and dynamic constraints. Journal of Computational Design


and Engineering, 4(2):106–130.
ED

Sedaghati, R., Suleman, A., and Tabarrok, B. (2002). Structural Optimiza-


tion with Frequency Constraints Using the Finite Element Force Method.
AIAA Journal, 40(2):382–388.
PT

Soh, C. K. and Yang, Y. (2000). Genetic Programming-Based Approach


for Structural Optimization. Journal of Computing in Civil Engineering,
14(1):31–37.
CE

Stolpe, M. (2016). Truss optimization with discrete design variables: a criti-


cal review. Structural and Multidisciplinary Optimization, 53(2):349–374.
AC

Wang, D., Zhang, W. H., and Jiang, J. S. (2004). Truss optimization on shape
and sizing with frequency constraints. AIAA Journal, 42(3):622–630.

Wu, S. J. and Chow, P. T. (1995). Steady-state genetic algorithms for discrete


optimization of trusses. Computers & Structures, 56(6):979–991.

42
ACCEPTED MANUSCRIPT

Xu, B., Jiang, J., Tong, W., and Wu, K. (2003). Topology group concept for
truss topology optimization with frequency constraints. Journal of Sound
and Vibration, 261(5):911–925.

Yang, J. and Soh, C. K. (1998). Discussion of “Genetic Algorithms-Based

T
Methodologies for Design Optimization of Trusses” by Jiaping Yang and
Chee-Kiong Soh. Journal of Structural Engineering, 124(8):980–981.

IP
Yang, X.-S. (2010). Engineering optimization: an introduction with meta-

CR
heuristic applications. John Wiley & Sons.

Yang, Y. and Soh, C. K. (2000). Fuzzy Logic Integrated Genetic Program-


ming for Optimization and Design. Journal of Computing in Civil Engi-
neering, 14(4):249–254.
US
Yang, Y. and Soh, C. K. (2002). Automated optimum design of structures us-
ing genetic programming. Computers & Structures, 80(18-19):1537–1546.
AN
M
ED
PT
CE
AC

43

Anda mungkin juga menyukai