# Evaluation of combined Pareto multiobjective differential evolution on tuneable problems.

1. INTRODUCTIONIn most practical decision-making problems, the presence of multiple objectives or multiple criteria is evident [1]. Due to the multi-criteria nature of most real-world problems, multiobjective optimization problems (MOOPs) are very common particularly in engineering and scientific designs and applications. MOOPs involve multiple often conflicting objectives, which are to be optimized simultaneously. There is no single optimal solution to this class of problems; rather, the solution consists of a group of alternative trade-off solutions called Pareto-optimal or non-inferior solutions which must be considered equivalent in the absence of specialized information concerning the relative importance of the objectives [2, 3].

Evolutionary algorithms (EAs) are population-based meta-heuristic optimization algorithms that use biology-inspired mechanisms like mutation, crossover, natural selection and survival of the fittest in order to refine a set of solution candidates iteratively. EAs have often performed well approximating solutions to all types of problems because they ideally do not make any assumption about the underlying fitness landscape [4, 5]. Since EAs deal with a group of candidate solutions, it seems natural to use them in MOOPs to find a group of optimal solutions. Indeed, the applications of evolutionary multi-objective algorithms (EMOAs) to the solution of real world MOOPs have been demonstrated and they have been found very efficient in solving these classes of problems [6-8]. The development of the theory of EMOA in recent years has spurred researches in the field of development of EMOAs for the solution of real-world problems. Over the past decades, several studies involving the extension of evolutionary algorithms to solve multi-objective numerical optimization problems have been reported [9-11].

Differential evolution (DE) is currently one of the most popular heuristics being used for solving single-objective optimization problems in continuous search spaces. Due to its reported successes on a myriad of problems, its use has been extended to other types of problem domains, including multi-objective optimization [6, 12]. In recent times, several researches extending the application of DE for finding solutions in the multi-objective problem domains have been reported in the literatures [13-16]. For instance Robic and Filipic [17] proposed three strategies of a Pareto-based differential evolution for multi-objective optimization (DEMO). The suggested strategies are DEMO/parent, DEMO/closest/dec and DEMO/closest/obj. In DEMO/parent, a child replaces the parent at the same index if it dominates that parent while in DEMO/closest/dec, the child replaces the closest parent to it in the decision space. DEMO/closest/obj replaces a parent with a child in the objective function space. The performances of the strategies of DEMO were compared with six other methods of EMOAs on five benchmark test beds. It was found that DEMO outperformed the other algorithms especially those based on other EAs such as genetic algorithm (GA) and evolution strategies (ES). It was thus concluded that DEMO may be adopted as an alternative for solving MOOPs.

In [18], a combined Pareto multi-objective differential evolution (CPMDE) algorithm is introduced. The algorithm combines Pareto selection procedures for multi-objective differential evolution to implement a novel selection scheme at each generation. The performance of CPMDE was evaluated using common difficult test problems obtained from multi-objective evolutionary computation literatures. The ability of the algorithm in solving unconstrained, constrained and real optimization problems was demonstrated and competitive results obtained from its application suggest that it is a good alternative for solving multiobjective optimization problems. However, [1] has argued that most of these test problems are not tuneable and it is difficult to establish the feature of an algorithm that has been tested. Based on this argument, the author presented a systematic procedure of designing test problems for unconstrained and constrained multi-objective evolutionary optimization and constructed a set of six difficult test problems. These problems have further been studied by researcher in the field [9, 10, 14, 17]. Motivated by the preceding argument, we further test CPMDE using five (continuous) tuneable unconstrained multi-objective test problems and one constrained test problem. Furthermore, we apply CPMDE to solve a real world engineering design problem. Results obtained herein further corroborate the efficacy of CPMDE in solving multi-objective optimization problems.

The remainder of this paper is structured as follows. In section 2 we present a brief description of the CPMDE algorithm. A more thorough description of CPMDE can be found in [18]. Section 3 present methodologies adopted in benchmarking CPMDE using tuneable unconstrained multi-objective test beds, a constrained test problem and an application to the solution of an engineering design problem, while section 4 presents the results of the benchmark and evaluation of CPMDE. The paper is concluded in section 5.

2. THE CPMDE ALGORITHM

In CPMDE, the combined population of trial and target solutions at the end of every iteration is checked for non-dominated solutions. Solutions that will proceed to the next generation are selected using a combined Pareto ranking and Pareto dominance selection scheme [6]. Diversity among solutions in the obtained non-dominated set is promoted using a harmonic average crowding distance measure [19]. Furthermore, boundary constraints are handled using the bounce-back strategy [12] while equality and inequality constraints are handled using the constrained-domination technique suggested by [1]. The CPMDE algorithm is summarized as follows [18]:

1. Input the required DE parameters like number of individuals in the population Np, mutation scaling factor F, crossover probability Cr, maximum number of iterations/generations gMax, number of objective functions k, number of decision variables/parameters D, upper and lower bounds of each variable, etc.

2. Initialize all solution vectors randomly within the limits of the variable bounds.

3. Set the generation counter, g = 0.

4. Generate a trial population of size Np using DE's mutation and crossover operations [12].

5. Perform a domination check on the combined trial and target population and mark all non-dominated solutions as "non-dominated" while marking others as "dominated".

6. Play domination tournament at each population index.

i. If the trial solution is marked "non-dominated" and the target is marked "dominated" then the trial vector replaces the target vector.

ii. If the trial solution is marked "dominated" and the target is marked "nondominated" then the trial vector is discarded.

iii. If both solutions are marked "dominated", then replace the target vector if it is dominated by the trial vector or if they are non-dominated with respect to each other.

iv. If both vectors are marked "non-dominated", then note down the index and proceed to the next index. When all solutions marked "non-dominated" from steps i-iii above are installed in the next generation, then sort out all solutions noted in step iv one at a time using the harmonic average crowding distance measure [19]. The solution with a greater harmonic average distance is selected to proceed to the next generation.

7. Increase the generation counter, g, by 1. i.e. g = g + 1.

8. If g < gMax, then go to step 4 above else go to step 9.

9. Remove the dominated solutions in the last generation.

10. Output the non-dominated solutions.

* Note domination checks are performed using the naive and slow method [1].

Source: [18]

3. EVALUATING AND BENCHMARKING CPMDE

The performance of CPMDE is compared with 13 state-of-the-art EMOAs on five unconstrained benchmark test beds and one constrained test problem. Furthermore, the performance of CPMDE on an engineering two-bar truss design problem is demonstrated. Other algorithms used in benchmarking CPMDE in this study include NSGA-II (real coded), NSGA-II (binary coded), SPEA, PAES, PDEA, MODE, MODE-E (MODE with external archive and crowding distance measure), MOPSO, SDE, DEMO/parent, DEMO/closest/dec, DEMO/closest/obj and MDEA.

3.1 Description of benchmark test problems

Five Zitzler-Deb-Thiele (ZDT1, ZDT2, ZDT3, ZDT4 and ZDT6) test problems, which are common tuneable difficult benchmark problems used in the literatures [9, 17, 20, 21] are chosen for evaluating the performance of CPMDE on unconstrained optimization problems. These are bi-objective problems in which both objectives are to be minimized. Each problem poses a different type of difficulty to EMOAs. The definitions and descriptions of all test functions are taken from literatures [1, 8, 14, 17] and summarized in Table I.

ZDT1 is a 30-variable problem having a convex Pareto-optimal front. The difficulty an EMOA may face on this problem is in tackling the large number of decision variables. ZDT2 is also a 30-variable problem but has a non-convex Pareto-optimal front. The non-convexity of the front is the major difficulty posed here. ZDT3 is a 30-variable problem having a number of disconnected Pareto-optimal fronts. Finding uniform spread of solutions on all discontinuous regions is the challenge in this problem. ZDT4 is a 10-variable problem with 219 local optimal fronts. Escaping all local non-dominated fronts to converge to the global optimal front is a real challenge in this problem. ZDT6 is a 10-variable problem having a nonconvex Pareto-optimal front with non-uniform distribution of solutions on the front. Finding a uniform spread of solution on this non-convex front poses a challenge to EMOAs.

To evaluate the performance of CPMDE on constrained optimization problems, the problem CONSTR is used [1, 8]. This is a bi-objective problem with two constraints. Both of the objectives are to be minimized. CONSTR has a convex Pareto-optimal front with two distinct segments. Finding uniform spread of solutions on both segments while satisfying both constraints is a challenge in this problem.

3.2 Performance measures

Performance measures that exist for EMOA evaluation are multifarious [3, 22]. However, in order to provide a uniform basis for comparison of EMOAs used in this study, the three performance measures reported in the published studies were adopted [9, 14, 17]. The generational distance and convergence metric are used to evaluate convergence to the global Pareto-optimal front while the diversity metric is employed to measure the spread of solutions on the obtained non-dominated front.

Generational distance

This is the average distance of the non-dominated set of solutions in a set Q from a set of chosen Pareto-optimal solutions. It is computed using eq. (1):

GD = [([[absolute value of Q].summation over (i=1)][d.sup.p.sub.i]).sup.1/p]/[absolute value of Q]

Forp = 2, [d.sub.i] is the Euclidean distance (in the objective space) between the solutions of Q and the nearest member in the true Pareto-front. An algorithm having a small value of GD is better [1].

Convergence metric

This is a special case of the generational distance where p =1. Convergence metric is computed using eq. (2) [17]:

Y = [[absolute value of Q].summation over (i=1)][d.sub.i]/[absolute value of Q] (2)

Diversity metric

This metric measures the extent and spread of solutions in the obtained non-dominated front. It is computed using eq. (3):

[DELTA] = [M.summation over (m=1)][d.sup.e.sub.m] + [[absolute value of Q]-1.summation over (i=1)][absolute value of [d.sub.i] - [bar.d]]/[M.summation over (m=1)][d.sup.e.sub.m] + ([absolute value of Q] - 1)[bar.d] (3)

where [d.sub.i] is the Euclidean distance (in the objective space) between consecutive solutions in the obtained non-dominated front Q, and [bar.d] is the average of these distances. The parameter [d.sup.e.sub.m] is the Euclidean distances between the extreme solutions of the Pareto front and the boundary solution of the obtained non-dominated front Q with respect to each objective m. An algorithm with a smaller value of diversity metric [DELTA] is better [1].

3.3 Experimental setup

In this study, DE/rand/1/bin variant of DE was used as the base for CPMDE. Cr and F were set at 0.3. Population size Np was set to 100 and the algorithm was run for a maximum number of generations, gMax = 250. A set of 500 uniformly spaced solutions were taken from the Pareto-optimal set for computation of all metrics. Averages and variances of metric values over 10 runs are reported in this study. For comparison of the performance of CPMDE with NSGA-II on the problem CONSTR, the following parameters are used: Cr = F = 0.3, Np = 40 and gMax = 200. Harmonic average crowding distances are computed using the two nearest neighbours on either sides of a solution.

3.4 Two-bar truss design problem

To demonstrate the applicability of CPMDE in solving real-world optimization problems, the algorithm was applied to design a two-bar truss system. A problem originally studied by [23] using the e-constraint method and further studied by [24] using NSGA-II is adopted here. A schematic representation of the two-bar truss is depicted in Fig. 1.

The truss is designed to carry a certain load without elastic failure. Thus, in addition to the objective of designing the truss for minimum volume (which is equivalent to designing for minimum cost of fabrication); there are additional objectives of minimizing stresses in each of the two members AC and BC. The two-objective constrained optimization problem for three decision variables y (vertical distance between B and C in m), [x.sub.1] (cross-sectional area of AC in [m.sup.2]) and [x.sub.2] (cross-sectional area of BC in [m.sup.2]) is formulated as follows [24]:

Objective function 1 (Minimize volume): [f.sub.1]([x.sub.1], [x.sub.2], y) = [x.sub.1][square root of 16 + [y.sup.2]] + [x.sub.2][square root of 1 + [y.sup.2]]

Objective function 2 (Minimize stress): [f.sub.2]([x.sub.1], [x.sub.2], y) = max([[sigma].sub.AC], [[sigma].sub.BC])

where: [[sigma].sub.AC] = 20[square root of 16 + [y.sup.2]]/[yx.sub.1], [[sigma].sub.BC] = 80[square root of 1 + [y.sup.2]]/[yx.sub.2]

Subject to: max([[sigma].sub.AC], [[sigma].sub.BC]) [less than or equal to] 1([10.sup.5])

Bound constraints: 1 [less than or equal to] y [less than or equal to] 3, 0 [less than or equal to] [x.sub.1], [x.sub.2] [less than or equal to] 0.01

[[sigma].sub.AC] and [[sigma].sub.BC] are the stresses in members AC and BC respectively. On this problem, the following settings are used for CPMDE: Cr = 0.9, F = 0.5, Np = 100 and gMax = 100.

4. RESULTS AND DISCUSSION

The mean and variance of the convergence metric on the unconstrained test beds, over 10 runs of CPMDE are reported in Table II while those of the diversity metric are presented in Table III. The convergence metric for PDEA was not available; therefore for comparative study on this algorithm, the available generational distance metric is extracted from literatures and presented along with those of DEMO and CPMDE in Table IV. Diversity metrics for MODE on all test functions are not available because they were not calculated in the original study. Also, the performance metric for MOPSO on the test problem ZDT4 is not available. The authors reported that this algorithm failed on this test bed.

Reported values of generational distances, convergence and diversity metrics for other algorithms used in benchmarking CPMDE are taken from correlative literatures [9, 14, 17, 20, 21] and presented in the respective tables. Best mean results are shown in boldface. Fig. 2 depicts the convergence of the obtained non-dominated front to the true Pareto-optimal front in problems ZDT1, ZDT2, ZDT3, ZDT4, ZDT6 and CONSTR respectively. The values of the test metrics are indicated on the respective plots. Fig. 3 shows the performance of NSGA-II and CPMDE, respectively, for 200 generations on the CONSTR problem. Fig. 4 shows the results obtained by NSGA-II and CPMDE on the two-bar truss design problem.

From the results in Tables II - IV, it is found that CPMDE outperformed all other algorithms on ZDT1 test bed as it produced the minimum values of all test metrics in all cases (Y = 0.000755, GD = 0.000086, [DELTA] = 0.241173). Binary coded NSGA-II performed second with respect to convergence (Y = 0.000894) while MDEA was the second best with respect to diversity ([DELTA] = 0.283708). MDEA and DEMO/parent showed slightly better convergence property on ZDT2, however, CPMDE performed best with respect to diversity preservation this test best reporting a value of [DELTA] = 0.266395. Therefore the performance of CPMDE on this problem is comparable to MDEA and DEMO but better than those of other algorithms. Table II shows that CPMDE performed best in converging to the Pareto-optimal front of ZDT3 with a convergence metric, Y = 0.000916. However, the reported diversity metric for MDEA and all versions of DEMO were better. Hence, the performance of CPMDE is comparable with MDEA and DEMO on this test bed. The advantage of CPMDE in converging to the global Pareto-optimal front in deceptive multi-modal functions is amply demonstrated on test problem ZDT4. Here, CPMDE outperformed all other algorithms in convergence and diversity (Y = 0.000731, GD = 0.000085, [DELTA] = 0.203378). The runner-ups in this case are DEMO/closest/dec (Y = 0.001016) on convergence and MODE-E on diversity ([DELTA] = 0.338330). The convergence metric on this problem is clearly smaller than those of other algorithms. On ZDT6, CPMDE performed second best with a convergence metric (Y = 0.000584) which is slightly smaller than 0.000584 reported for MDEA. However, CPMDE produced the best diversity ([DELTA] = 0.217533) on this problem. On all unconstrained problems, CPMDE produces variance values of zero for convergence metrics and generational distances (Tables II and IV). This suggests that CPMDE is reliable and stable in converging to the true Pareto front on these test beds.

By inspection, Fig. 3 shows that while NSGA-II was able to produce solutions covering roughly 80 % of the Pareto-optimal front for 200 iterations on the CONSTR test problem, CPMDE spanned the entire front with better convergence property. Therefore, CPMDE outperforms NSGA-II which is one of the state-of-the-art algorithms on this problem.

On the two-bar truss design problem, e-constraint found only five solutions with spread (0.004445 [m.sup.3], 89983 kPa) - (0.004833 [m.sup.3], 83268 kPa) while NSGA-II found many solutions in the range (0.00407 [m.sup.3], 99755 kPa) - (0.05304 [m.sup.3], 8439 kPa). CPMDE was also able to find many solutions spanning the range (0.00408 [m.sup.3], 98787 kPa) - (0.07384 [m.sup.3], 8433 kPa). If minimization of stress is important, NSGA-II finds a solution with stress as low as 8439 kPa, whereas the e-constraint method found a solution with minimum stress of 83268 kPa, an order of magnitude higher than that found in NSGA-II [9]. CPMDE found a solution with minimum stress of 8433 kPa which is slightly less than that found by NSGA-II, thus, the performance of CPMDE is comparable with NSGA-II on this problem. CPMDE produces many quality nondominated solutions on the Pareto-optimal front of this problem in a single run (Fig. 4). This shows that CPMDE can perform well on real-world engineering problems.

5. CONCLUSION

A benchmark of Combined Pareto multi-objective differential evolution (CPMDE) on tuneable multi-objective test problems is presented in this study. The ability of CPMDE in solving constrained and real multi-objective optimization problems was also illustrated. The ability of CPMDE to converge to the global Pareto-optimal front in deceptive multi-modal functions is amply demonstrated on test problem ZDT4 which has 21 billion local optimal fronts. Among the 14 algorithms compared in this study, CPMDE produced the best convergence in 3 out of the 5, and best diversity in 4 out of 5 unconstrained test beds. Also, the variances of the metrics suggest that the algorithm is stable on the test beds. Furthermore, CPMDE was applied to solve a real-world problem where it's efficacy on such problems was confirmed. Competitive results obtained from the application of CPMDE suggest that it is a good alternative for solving multi-objective optimization problems. Therefore, this study further corroborates that CPMDE is adoptable as a method of EMOA for solving real-world MOOPs.

DOI: 10.2507/IJSIMM13(3)2.264

REFERENCES

[1] Deb, K. (2001). Multi-objective optimization using evolutionary algorithms, Wiley, Chichester

[2] Deb, K. (2011). Multi-objective optimization using evolutionary algorithms: An introduction, KanGAL Report Number 2011003, Vol. 1, 22-45

[3] Zhou, A.; Qu, B.-Y.; Li, H.; Zhao, S.-Z.; Suganthan, P. N.; Zhang, Q. (2011). Multiobjective evolutionary algorithms: A survey of the state of the art, Swarm and Evolutionary Computation, Vol. 1, No. 1, 32-49, doi:10.1016/j.swevo.2011.03.001

[4] Olofintoye, O.; Adeyemo, J.; Otieno, F. (2013). Evolutionary algorithms and water resources optimization, Schutze, O.; Coello Coello, C. A.; Tantar, A.-A.; Tantar, E.; Bouvry, P.; Moral, P. D.; Legrand, P. (Eds.), EVOLVE - A Bridge between Probability, Set Oriented Numerics, and Evolutionary Computation II, Vol. 175, Springer-Verlag, Berlin, 491-506

[5] Weise, T. Global Optimization Algorithms - Theory and Application (book), from http://www.itweise.de/projects/book.pdf, accessed on 29-01-2014

[6] Mezura-Montes, E.; Reyes-Sierra, M.; Coello Coello, C. A. (2008). Multi-objective optimization using differential evolution: a survey of the state-of-the-art, Chakraborty, U. K. (Ed.), Advances in differential evolution, Vol. 143, Springer-Verlag, Berlin, 173-196

[7] Qin, H.; Zhou, J.; Lu, Y.; Wang, Y.; Zhang, Y. (2010). Multi-objective differential evolution with adaptive Cauchy mutation for short-term multi-objective optimal hydro-thermal scheduling, Energy Conversion and Management, Vol. 51, No. 4, 788-794, doi:10.1016/j.enconman.2009. 10.036

[8] Reddy, M. J.; Kumar, D. N. (2007). Multiobjective differential evolution with application to reservoir system optimization, Journal of Computing in Civil Engineering, Vol. 21, No. 2, 136146, doi:10.1061/(ASCE)0887-3801(2007)21:2(136)

[9] Deb, K.; Pratap, A.; Agarwal, S.; Meyarivan, T. (2002). A fast and elitist multiobjective genetic algorithm: NSGA-II, IEEE Transactions on Evolutionary Computation, Vol. 6, No. 2, 182-197, doi:10.1109/4235.996017

[10] Fonseca, C. M.; Fleming, P. J. (1993). Genetic algorithms for multiobjective optimization: formulation, discussion and generalization, Proceedings of the Fifth International Conference on Genetic Algorithms, 415-423

[11] Knowles, J.; Corne, D. (1999). The Pareto archived evolution strategy: a new baseline algorithm for Pareto multiobjective optimisation, Proceedings of the 1999 IEEE Congress on Evolutionary Computation, Vol. 1, 98-105

[12] Price, K. V.; Storn, R. M.; Lampinen, J. A. (2005). Differential Evolution: A Practical Approach to Global Optimization, Springer-Verlag, Berlin

[13] Abbass, H. A.; Sarker, R. (2002). The Pareto differential evolution algorithm, International Journal on Artificial Intelligence Tools, Vol. 11, No. 4, 531-552, doi: 10.1142/ S0218213002001039

[14] Adeyemo, J. A.; Otieno, F. A. O. (2009). Multi-objective differential evolution algorithm for solving engineering problems, Journal of Applied Sciences, Vol. 9, No. 20, 3652-3661, doi: 10.3923/jas.2009.3652.3661

[15] Babu, B. V.; Jehan, M. M. L. (2003). Differential evolution for multi-objective optimization, Proceedings of the 2003 IEEE Congress on Evolutionary Computation, Vol. 4, 2696-2703

[16] Ali, M.; Siarry, P.; Pant, M. (2012). An efficient Differential Evolution based algorithm for solving multi-objective optimization problems, European Journal of Operational Research, Vol. 217, No. 2, 404-416, doi:10.1016/i.eior.2011.09.025

[17] Robic, T.; Filipic, B. (2005). DEMO: Differential evolution for multiobjective optimization, Coello Coello, C. A.; Aguirre, A. H.; Zitzler, E. (Eds.), Evolutionary Multi-Criterion Optimization, Vol. 3410, Springer-Verlag, Berlin, 520-533

[18] Olofintoye, O.; Adeyemo, J.; Otieno, F. (2014). A combined Pareto differential evolution approach for multi-objective optimization, Schutze, O.; Coello Coello, C. A.; Tantar, A.-A.; Tantar, E.; Bouvry, P.; Moral, P. D.; Legrand, P. (Eds.), EVOLVE--A Bridge between Probability, Set Oriented Numerics, and Evolutionary Computation III, Vol. 500, Springer International Publishing, 213-231

[19] Huang, V. L.; Suganthan, P. N.; Qin, A. K.; Baskar, S. (2005). Multi-objective differential evolution with external archive and harmonic distance-based diversity measure, Nanyang Technological University, Technical Report, 1-25

[20] Madavan, N. K. (2002). Multiobjective optimization using a Pareto differential evolution approach, Proceedings of the 2002 IEEE Congress on Evolutionary Computation, 1145-1150

[21] Xue, F.; Sanderson, A. C.; Graves, R. J. (2003). Pareto-based multi-objective differential evolution, Proceedings of the 2003 IEEE Congress on Evolutionary Computation, Vol. 2, 862869

[22] Schutze, O.; Esquivel, X.; Lara, A.; Coello Coello, C. A. (2012). Using the averaged Hausdorff distance as a performance measure in evolutionary multiobjective optimization, IEEE Transactions on Evolutionary Computation, Vol. 16, No. 4, 504-522, doi: 10.1109/TEVC. 2011.2161872

[23] Palli, N.; Azarm, S.; McCluskey, P.; Sundararajan, R. (1998). An interactive multistage sinequality constraint method for multiple objectives decision making, Journal of Mechanical Design, Vol. 120, No. 4, 678-686, doi:10.1115/1.2829331

[24] Deb, K.; Pratap, A.; Moitra, S. (2000). Mechanical component design for multiple objectives using elitist non-dominated sorting GA, Schoenauer, M.; Deb, K.; Rudolph, G.; Yao, X.; Lutton, E.; Merelo, J. J.; Schwefel, H.-P. (Eds.), Parallel Problem Solving from Nature PPSN VI, Vol. 1917, Springer-Verlag, Berlin, 859-868

Adeyemo, J. A. & Olofintoye, O. O. *

Department of Civil Engineering and Surveying, Durban University of Technology, P.O. Box 1334, Durban, 4000, South Africa

E-Mail: josiaha@dut.ac.za, olofintoye.oo@gmail.com (Corresponding author)

Table I: Summary of benchmark test problems. Problem n Variable bounds Objective functions and constraints ZDT1 30 [0, 1] [f.sub.1](x) = [x.sub.1] [f.sub.2](x) = g(x)(1 - [square root of [f.sub.1](x)/g(x)]) g(x) = 1 + 9([n.summation over (i=2)][x.sub.i])/ (n - 1) ZDT2 30 [0, 1] [f.sub.1](x) = [x.sub.1] [f.sub.2](x) = g(x)(1 - [([f.sub.1](x)/ g(x)).sup.2]) g(x) = 1 + 9([n.summation over (i=2)][x.sub.i])/ (n - 1) ZDT3 30 [0, 1] [f.sub.1](x) = 1 [f.sub.2](x) = g(x)[1 - [[square root of [x.sub.1]/g(x)]] - [[x.sub.1]/g(x)]sin (10[pi][x.sub.1])] g(x) = 1 + 9([n.summation over (i=2)][x.sub.i])/ (n - 1) ZDT4 10 [x.sub.1] [member of] [f.sub.1](x) = [x.sub.1] [0, 1] [x.sub.i] [member of] [f.sub.2](x) = g(x)[1 - [-5, 5] [square root of ([x.sub.1]/g(x))]] i = 2, 3, ..., n g(x) = 1 + 10(n - 1) + [summation over (i=2)] ([x.sup.2.sub.i] - 10cos(4[pi][x.sub.i])) ZDT6 10 [0, 1] [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] [f.sub.2](x) = g(x)[1 - [([f.sub.1](x)/ g(x)).sup.2]] g(x) = 1 + [9/n - 1] [n.summation over (i=2)][x.sub.i] CONSTR 10 [x.sub.1] [member of] [f.sub.1](x) = [x.sub.1] [0.1, 1] [x.sub.2] [member of] [f.sub.2](x) = 1 + [0, 5] [x.sub.2]/[x.sub.1] Subject to: [g.sub.1](x) = [x.sub.2] + 9[x.sub.1] [greater than or equal to] 6 [g.sub.2](x) = -[x.sub.2] + 9[x.sub.1] [greater than or equal to] 1 Problem Pareto-optimal Comments solutions ZDT1 [x.sub.1] [member of] Convex [0,1] [x.sub.i] = 0 i = 2, 3, ..., n ZDT2 [x.sub.1] [member of] Non-convex [0,1] [x.sub.i] = 0 i = 2, 3, ..., n ZDT3 [x.sub.1] [member of] Convex, [0,1] Discontinuous Not [x.sub.i] = 0 all points in 0 [less than or equal to] [x.sub.l] [less than or i = 2, 3, ..., n equal to] 1 lie on the Pareto- optimal front ZDT4 [x.sub.1] [member of] Convex, Deceptive, [0,1] Multiple local, [x.sub.i] = 0 Optimal fronts i = 2, 3, ..., n ZDT6 [x.sub.1] [member of] Nonconvex, [0,1] Nonuniformly spaced solutions on Pareto front, Low density of solutions near Pareto front [x.sub.i] = 0 i = 2, 3, ..., n CONSTR 0.39 [less than or Convex, equal to] [x.sub.1] Constrained, [less than or Segmented Pareto equal to] 0.67: front [x.sub.2] = 6 - 9[x.sub.1] and 0.67 [less than or equal to] [x.sub.1] [less than or equal to] 1.00: [x.sub.2] = 0 Table II: Convergence metrics on unconstrained test beds. Algorithm ZDT1 NSGA-II 0.033482 [+ or -] 0.004750 (real coded) NSGA-II 0.000894 [+ or -] 0.000000 (binary coded) SPEA 0.001799 [+ or -] 0.000001 PAES 0.082085 [+ or -] 0.008679 PDEA N/A MODE 0.005800 [+ or -] 0.000000 MODE-E 0.001999 [+ or -] 0.000000 MOPSO 0.019659 [+ or -] 0.000012 SDE 0.002741 [+ or -] 0.000385 DEMO/parent 0.001083 [+ or -] 0.000113 DEMO/closest/dec 0.001113 [+ or -] 0.000134 DEMO/closest/obj 0.001132 [+ or -] 0.000136 MDEA 0.000921 [+ or -] 0.000005 CPMDE 0.000755 [+ or -] 0.000000 Algorithm ZDT2 NSGA-II 0.072391 [+ or -] 0.031689 (real coded) NSGA-II 0.000824 [+ or -] 0.000000 (binary coded) SPEA 0.001339 [+ or -] 0.000000 PAES 0.126276 [+ or -] 0.036877 PDEA N/A MODE 0.005500 [+ or -] 0.000000 MODE-E 0.001554 [+ or -] 0.000000 MOPSO 0.017093 [+ or -] 0.000133 SDE 0.002203 [+ or -] 0.000297 DEMO/parent 0.000755 [+ or -] 0.000045 DEMO/closest/dec 0.000820 [+ or -] 0.000042 DEMO/closest/obj 0.000780 [+ or -] 0.000035 MDEA 0.000640 [+ or -] 0.000000 CPMDE 0.000775 [+ or -] 0.000000 Algorithm ZDT3 NSGA-II 0.114500 [+ or -] 0.004940 (real coded) NSGA-II 0.043411 [+ or -] 0.000042 (binary coded) SPEA 0.047517 [+ or -] 0.000047 PAES 0.023872 [+ or -] 0.000010 PDEA N/A MODE 0.021560 [+ or -] 0.000000 MODE-E 0.002642 [+ or -] 0.000000 MOPSO 0.030469 [+ or -] 0.000067 SDE 0.002741 [+ or -] 0.000120 DEMO/parent 0.001178 [+ or -] 0.000059 DEMO/closest/dec 0.001197 [+ or -] 0.000091 DEMO/closest/obj 0.001236 [+ or -] 0.000091 MDEA 0.001139 [+ or -] 0.000024 CPMDE 0.000916 [+ or -] 0.000000 Algorithm ZDT4 NSGA-II 0.513053 [+ or -] 0.118460 (real coded) NSGA-II 3.227636 [+ or -] 7.307630 (binary coded) SPEA 7.340299 [+ or -] 6.572516 PAES 0.854816 [+ or -] 0.527238 PDEA N/A MODE 0.638950 [+ or -] 0.500200 MODE-E 0.030689 [+ or -] 0.004867 MOPSO N/A SDE 0.100100 [+ or -] 0.446200 DEMO/parent 0.001037 [+ or -] 0.000134 DEMO/closest/dec 0.001016 [+ or -] 0.000091 DEMO/closest/obj 0.041012 [+ or -] 0.063920 MDEA 0.048962 [+ or -] 0.536358 CPMDE 0.000731 [+ or -] 0.000000 Algorithm ZDT6 NSGA-II 0.296564 [+ or -] 0.013135 (real coded) NSGA-II 7.806798 [+ or -] 0.001667 (binary coded) SPEA 0.221138 [+ or -] 0.000449 PAES 0.085469 [+ or -] 0.006664 PDEA N/A MODE 0.026230 [+ or -] 0.000861 MODE-E 0.005998 [+ or -] 0.000005 MOPSO 0.751692 [+ or -] 0.151000 SDE 0.000624 [+ or -] 0.000060 DEMO/parent 0.000629 [+ or -] 0.000044 DEMO/closest/dec 0.000630 [+ or -] 0.000021 DEMO/closest/obj 0.000642 [+ or -] 0.000029 MDEA 0.000436 [+ or -] 0.000055 CPMDE 0.000584 [+ or -] 0.000000 Table III: Diversity metrics on unconstrained test beds. Algorithm ZDT1 NSGA-II 0.390307 [+ or -] 0.001876 (real coded) NSGA-II 0.463292 [+ or -] 0.041622 (binary coded) SPEA 0.784525 [+ or -] 0.004440 PAES 1.229794 [+ or -] 0.000742 PDEA 0.298567 [+ or -] 0.000742 MODE N/A MODE-E 0.306235 [+ or -] 0.001130 MOPSO 0.586728 [+ or -] 0.001480 SDE 0.382890 [+ or -] 0.001435 DEMO/parent 0.325237 [+ or -] 0.030249 DEMO/closest/dec 0.319230 [+ or -] 0.031350 DEMO/closest/obj 0.306770 [+ or -] 0.025465 MDEA 0.283708 [+ or -] 0.002938 CPMDE 0.241173 [+ or -] 0.000077 Algorithm ZDT2 NSGA-II 0.430776 [+ or -] 0.004721 (real coded) NSGA-II 0.435112 [+ or -] 0.024607 (binary coded) SPEA 0.755184 [+ or -] 0.004521 PAES 1.165942 [+ or -] 0.007682 PDEA 0.317958 [+ or -] 0.001389 MODE N/A MODE-E 0.298449 [+ or -] 0.000580 MOPSO 0.689580 [+ or -] 0.038200 SDE 0.345780 [+ or -] 0.003900 DEMO/parent 0.329151 [+ or -] 0.032408 DEMO/closest/dec 0.335178 [+ or -] 0.016985 DEMO/closest/obj 0.326821 [+ or -] 0.021083 MDEA 0.450482 [+ or -] 0.004211 CPMDE 0.266395 [+ or -] 0.000379 Algorithm ZDT3 NSGA-II 0.738540 [+ or -] 0.019706 (real coded) NSGA-II 0.575606 [+ or -] 0.005078 (binary coded) SPEA 0.672938 [+ or -] 0.003587 PAES 0.789920 [+ or -] 0.001653 PDEA 0.623812 [+ or -] 0.000225 MODE N/A MODE-E 0.504275 [+ or -] 0.000200 MOPSO 0.594200 [+ or -] 0.001150 SDE 0.525770 [+ or -] 0.043030 DEMO/parent 0.309436 [+ or -] 0.018603 DEMO/closest/dec 0.324934 [+ or -] 0.029648 DEMO/closest/obj 0.328873 [+ or -] 0.019142 MDEA 0.299354 [+ or -] 0.023309 CPMDE 0.353734 [+ or -] 0.000522 Algorithm ZDT4 NSGA-II 0.702612 [+ or -] 0.064648 (real coded) NSGA-II 0.479475 [+ or -] 0.009841 (binary coded) SPEA 0.798463 [+ or -] 0.014616 PAES 0.870458 [+ or -] 0.101399 PDEA 0.840852 [+ or -] 0.035741 MODE N/A MODE-E 0.338330 [+ or -] 0.003676 MOPSO N/A SDE 0.436300 [+ or -] 0.110000 DEMO/parent 0.359905 [+ or -] 0.037672 DEMO/closest/dec 0.359600 [+ or -] 0.026977 DEMO/closest/obj 0.407225 [+ or -] 0.094851 MDEA 0.406382 [+ or -] 0.062308 CPMDE 0.203378 [+ or -] 0.000400 Algorithm ZDT6 NSGA-II 0.668025 [+ or -] 0.009923 (real coded) NSGA-II 0.644477 [+ or -] 0.035042 (binary coded) SPEA 0.849389 [+ or -] 0.002713 PAES 1.153052 [+ or -] 0.003916 PDEA 0.473074 [+ or -] 0.021721 MODE N/A MODE-E 0.335594 [+ or -] 0.019000 MOPSO 0.935686 [+ or -] 0.018500 SDE 0.361100 [+ or -] 0.036100 DEMO/parent 0.442308 [+ or -] 0.039255 DEMO/closest/dec 0.461174 [+ or -] 0.035289 DEMO/closest/obj 0.458641 [+ or -] 0.031362 MDEA 0.305245 [+ or -] 0.019407 CPMDE 0.217533 [+ or -] 0.000234 Table IV: Generational distance metrics on unconstrained test beds. Algorithm ZDT1 ZDT2 ZDT3 ZDT4 ZDT6 PDEA 0.000615 0.000652 0.000563 0.618258 0.023886 [+ or -] [+ or -] [+ or -] [+ or -] [+ or -] 0.000000 0.000000 0.000000 0.826881 0.003294 DEMO/parent 0.000230 0.000091 0.000156 0.000202 0.000074 DEMO [+ or -] [+ or -] [+ or -] [+ or -] [+ or -] 0.000048 0.000004 0.000007 0.000053 0.000004 /closest/dec 0.000242 0.000097 0.000162 0.000179 0.000075 DEMO [+ or -] [+ or -] [+ or -] [+ or -] [+ or -] 0.000028 0.000004 0.000013 0.000048 0.000002 /closest/obj 0.000243 0.000092 0.000169 0.004262 0.000076 [+ or -] [+ or -] [+ or -] [+ or -] [+ or -] 0.000050 0.000004 0.000017 0.006545 0.000003 CPMDE 0.000086 0.000087 0.000107 0.000085 0.000068 [+ or -] [+ or -] [+ or -] [+ or -] [+ or -] 0.000000 0.000000 0.000000 0.000000 0.000000

Printer friendly Cite/link Email Feedback | |

Title Annotation: | Original scientific paper |
---|---|

Author: | Adeyemo, J.A.; Olofintoye, O.O. |

Publication: | International Journal of Simulation Modelling |

Article Type: | Report |

Geographic Code: | 1USA |

Date: | Sep 1, 2014 |

Words: | 5687 |

Previous Article: | Inventory control system based on control charts to improve supply chain performances. |

Next Article: | Product attribute design using an agent-based simulation of an artificial market. |

Topics: |