You are on page 1of 3

Discussions and Closures

precise to compare different optimization algorithms based on the


Discussion of “Performance of Shuffled
number of simulations (function evaluations) instead of using
Frog-Leaping Algorithm in Finance-Based the number of iterations as the performance criterion.
Scheduling” by Anas Alghazi, The number of function evaluations is an appropriate criterion to
Shokri Z. Selim, and Ashraf Elazouni compare different meta-heuristic algorithms and evolutionary
methods; this has been widely used to compare static and dynamic
May/June 2012, Vol. 26, No. 3, pp. 396–408.
penalty functions (Bozorg Haddad and Mariño 2007), to compare
DOI: 10.1061/(ASCE)CP.1943-5487.0000157
two different algorithms (GA and honey-bee mating optimization)
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/29/18. Copyright ASCE. For personal use only; all rights reserved.

(Bozorg Haddad et al. 2009), and to compare two approaches of


E. Fallah-Mehdipour 1; O. Bozorg Haddad 2; and
one algorithm (continuous and discrete particle swarm optimiza-
M. A. Mariño, Dist.M.ASCE 3
1
tion) (Noory et al. 2012).
Ph.D. Candidate, Dept. of Irrigation and Reclamation Engineering, In population-based algorithms such as GA and SFLA, the num-
Faculty of Agricultural Engineering and Technology, College of Agri-
ber of function evaluations is equal to the population size × number
culture and Natural Resources, Univ. of Tehran, Karaj, Tehran, Iran.
E-mail: Falah@ut.ac.ir of iterations. In contrast, in a single point-based algorithm structure
2
Associate Professor, Dept. of Irrigation and Reclamation Engineering, such as SA, the number of iterations is equal to the number of func-
Faculty of Agricultural Engineering and Technology, College of Agri- tion evaluations. A comparison of meta-heuristic algorithms based
culture and Natural Resources, Univ. of Tehran, Karaj, Tehran, Iran on the number of function evaluations is independent of the
(corresponding author). E-mail: OBHaddad@ut.ac.ir processing time and computer capabilities, even when using differ-
3
Distinguished Professor Emeritus, Dept. of Land, Air and Water ent computers, especially in discrete and complex decision space
Resources, Dept. of Civil and Environmental Engineering, Dept. of problems such as finance-based problems.
Biological and Agricultural Engineering, Univ. of California, In this discussion, to present a comparison of SA, GA, and
139 Veihmeyer Hall, Davis, CA 95616-8628. E-mail: MAMarino@ SFLA based on function evaluation and time processing, a
ucdavis.edu
mathematical benchmark problem by Gen and Cheng (1997) is
considered
The authors used the shuffled frog-leaping algorithm (SFLA) to Max:fðx1 ; x2 Þ ¼ 21.5 þ x1 sinð4πx1 Þ þ x2 sinð20πx2 Þ ð1Þ
solve two large benchmark finance-based problems with 120
and 210 activities, respectively. In the first step, the performances
of simulated annealing (SA) and SFLA employing the approaches −3.0 ≤ x1 ≤ 12.1 ð2Þ
of the repaired and penalized solutions were compared by using a
laptop with a 2 GHz processor and 1 GB of random-access memory 4.1 ≤ x2 ≤ 5.8 ð3Þ
(RAM). In the second step, the performance of the SFLA was com-
pared against two common meta-heuristic algorithms: (1) genetic where f = objective function and x1 and x2 = first and second
algorithm (GA); and (2) SA. In this step, they used an Intel i7 quad decision variables, respectively. Table 1 shows the values of param-
core processor with 4 GB RAM and a 2 GHz processor with 1 GB eters used for SA, GA, and SFLA.
RAM for the problems with 120 and 210 activities, respectively. As shown, the number of function evaluations is 25,000 for all
The authors reported the profit of the project as the objective aforementioned algorithms. Because of the random base of meta-
function against the processing time. A comparison of the conver- heuristic algorithms, 10 different runs were conducted for each
gence of the different algorithms was presented in Figs. 6 and 7 of optimization algorithm. The best (maximum) results of SA, GA,
the original paper. Although the authors tried to compare different and SFLA among 10 runs are 38.3505, 38.829, and 38.850, respec-
algorithms under the same conditions (such as the same population tively. Although there is no considerable difference between differ-
size and number of iterations) for GA and SFLA, the comparison ent results of these algorithms, SA, GA, and SFLA yielded the best
directly depends on the capabilities of the computer, especially the results in 2.10, 1.67, and 0.13 s, respectively.
core processor and RAM. It is possible to change results from one Fig. 1 shows the convergence of the aforementioned algorithms
computer to another, or to use other parameters for each algorithm. against the number of function evaluations.
Random-based iterative algorithms have different structures and As shown, in the primary function evaluations, the SA yields a
consist of various elements. Thus, in each iteration of each algo- better (higher) objective function value than the SFLA. However,
rithm, a unique procedure occurs that differs from the others. In after approximately 4,050 function evaluations, the SFLA improves
each iteration, many simulations of the problems take place; after the obtained solution. This comparison is not directly related to
each simulation for a candidate solution, the fitness function is the structure of the algorithm or the capability of the computer,
calculated to evaluate the suitability of such solution. In many and is more useful and applicable when comparing meta-heuristic
optimization problems, especially in engineering problems, the
simulation procedure is much more time-consuming than the opti- Table 1. Parameter Values for SA, GA, and SFLA
mization procedure. Thus, optimization algorithms, even with the
same number of iterations, may not require the same computational SA GA SFLA
time; adversely, in the same processing time, different algorithms Iteration 25,000 Population size 500 Population size 500
may not end up with the same number of iterations. Cooling factor 0.955 Generation 50 Iteration 50
On the other hand, the number of simulations (function evalu- Boltzmann’s 1 Crossover 0.5 Memeplex size 10
ations) does not depend on hardware capabilities or expertise in constant probability
programming, whereas the number of iterations highly depends Initial 0.5 Mutation 0.1 Submemeplex 5
temperature probability size
on the aforementioned shortcomings. However, it would be more

JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / MARCH/APRIL 2014 / 327

J. Comput. Civ. Eng., 2014, 28(2): 328-329


40

35

Objective function 30

25
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/29/18. Copyright ASCE. For personal use only; all rights reserved.

20

GA SFLA SA
15

10

11050
50

1050

2050

3050

4050

5050

6050

7050

8050

9050

10050

12050

13050

14050

15050

16050

17050

18050

19050

20050

21050

22050

23050

24050
Function evaluation

Fig. 1. Meta-heuristic convergence for the same function evaluation

algorithms. Thus, it is recommended to assess the performance in the implementation of the genetic algorithm (GA). Alghazi
of SFLA in finance-based scheduling problems by using function (2009) proved that the GA with repaired chromosomes was
evaluations. more efficient than the GA with penalized chromosomes. In con-
trast, the implementations of simulated annealing (SA) and the
shuffled frog-leaping algorithm (SFLA) in the current paper
References proved that the approach of assigning penalties to the infeasible
Bozorg Haddad, O., Afshar, A., and Mariño, M. A. (2009). “Optimization schedules was more efficient than the approach of repairing
of non-convex water resource problems by honey-bee mating optimi- the infeasible solutions. This finding was reached based on
zation (HBMO) algorithm.” Eng. Comput., 26(3), 267–280. the 210 activity network, under a constrained credit limit
Bozorg Haddad, O., and Mariño, M. A. (2007). “Dynamic penalty function of $60,000, and using a laptop with a 2 GHz processor and
as a strategy in solving water resources combinatorial optimization 1 GB of random access memory (RAM). Accordingly, the repair
problems with honey-bee mating optimization (HBMO) algorithm.” approach was followed in the implementation of the GA, whereas
J. Hydroinform., 9(3), 233–250. the penalty approach was followed in the implementations of
Gen, M., and Cheng, R. W. (1997). Genetic algorithms and engineering
the SA and SFLA to compare the performance of the three
design, Wiley, New York.
Noory, H., Liaghat, A. M., Parsinejad, M., and Bozorg Haddad, O. (2012). meta-heuristics.
“Optimizing irrigation water allocation and multicrop planning using Initially, the comparison was conducted based on the 210
discrete PSO algorithm.” J. Irrig. Drain. Eng., 10.1061/(ASCE)IR activity network using a laptop with a 2 GHz processor and
.1943-4774.0000426, 437–444. 1 GB RAM. Upon completion of the first round of the manuscript
revision, the reviewers requested a comparison between the heu-
ristics using other networks; therefore, the 120 activity network
was used. When the authors were working on the implementations
Closure to “Performance of Shuffled required to compare the performance between the three meta-heu-
Frog-Leaping Algorithm in Finance-Based ristics, a more powerful laptop with an Intel i7 quad core processor
Scheduling” by Anas Alghazi, and 4 GB RAM was made available; thus, it was utilized for these
Shokri Z. Selim, and Ashraf Elazouni implementations. The authors reported the performance criterion
of the processing time against the objective function of the project
May/June 2012, Vol. 26, No. 3, pp. 396–408.
profit.
DOI: 10.1061/(ASCE)CP.1943-5487.0000157 Although the processing time was utilized as the performance
criterion in the comparison between the GA, SA, and SFLA, this
Ashraf Elazouni 1
1 author agrees with the argument made in the discussion that the
Associate Professor, Construction Engineering and Management Dept.,
function evaluations are more applicable than the processing time
King Fahd Univ. of Petroleum and Minerals, P.O. Box 346, Dhahran
31261, Saudi Arabia. E-mail: elazouni@kfupm.edu.sa when comparing the meta-heuristic algorithms. The primary merits
of comparing by using the function evaluations is the elimination
of irrelevant factors, including the capability of the computer,
algorithm structure, and inefficient programming. However, ac-
Alghazi (2009) addressed the problem of generating infeasible cording to the mathematical benchmark problem presented in
schedules caused by the violation of the credit limit that arises the discussion, the time the best results were obtained by the

328 / JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / MARCH/APRIL 2014

J. Comput. Civ. Eng., 2014, 28(2): 328-329


meta-heuristics under comparison has yet to be reported along with References
specifications of the computer; this may give the user additional
criteria to evaluate the performance, particularly when the differ- Alghazi, A. (2009). “Finance-based scheduling of activity networks.”
ences between the objective functions of the meta-heuristics are M.Sc. thesis, System Engineering Department, King Fahd Univ. of
not significant. Petroleum and Minerals, Dhahran, Saudi Arabia.
Downloaded from ascelibrary.org by Universidad Nacional De Ingenieria on 10/29/18. Copyright ASCE. For personal use only; all rights reserved.

JOURNAL OF COMPUTING IN CIVIL ENGINEERING © ASCE / MARCH/APRIL 2014 / 329

J. Comput. Civ. Eng., 2014, 28(2): 328-329

You might also like