You are on page 1of 4

I.

De Falco

A Comparative Analysis of Evolutionary Algorithms for Function Optimisation


Research Institute on Parallel Information Systems National Research Council of Italy via P. Castellino 111, Naples, Italy, 80131 fivan, renato, dean, ernestog@irsip.na.cnr.it

R. Del Balio

A. Della Cioppa

E. Tarantino

without roulette wheel selection, so that they generate a new population of N ? 1 elements. The former best element is then inserted in this new population, and the cycle of life continues. By doing so, the best elements are treated as `super-individuals' and mated together, hoping that this can lead to a tter population. These concepts are taken from other sciences and mimic animal breeding. Another very important feature of BGAs is the fact that they represent solutions as sequences of real numbers, I. Introduction rather than sequences of bits or integers, so they allow A wide variety of heuristic techniques has been proposed a representation of continuous problems much better than to solve di erent kinds of optimisation problems. Among normal GAs. A BGA can be formally described by: these, the Genetic Algorithms (GAs) 1; 2] and their parallel versions 3] should be mentioned. They have BGA = (P0 ; N ; T ; ?; ; A; ) proved to be able to e ciently solve di cult optimisation problems. Recently a novel technique, the Breeder Genetic where P0 is the initial random population, N the Algorithms (BGAs) 4; 5], particularly suitable to deal with population size, T the truncation threshold, ? the continuous optimisation parameters, has been introduced. recombination operator, the mutation operator, A the BGAs are based on the evolution model typical of GAs, tness function and the termination criterion. Typical recombination operators are the Discrete however they di er from the `traditional' GAs in some concepts borrowed from population genetics like selection Recombination (DR), the Extended Line Recombination intensity and response to selection. Besides new genetic (ELR) and the Extended Intermediate Recombination operators capable of dealing with continuous variables (EIR) 5]. DR mixes the alleles in the sequences and have been de ned. Since it is reasonable to assume that can only generate corners in the hypercube de ned by the structural changes in a genetic approach will a ect its sequences, whereas ELR and EIR generate a new allele ability at solving problems, an appropriate set of test by means of a linear combination of the corresponding functions has been taken from the literature in the attempt alleles and can generate respectively any point on the lines to establish the e ectiveness of this new technique. The connecting the corresponding vertices and in the volume of performance of BGAs is compared with that obtained by the hypercube de ned by the sequences. Typical mutation means of di erent versions of existing serial and parallel operators are discrete and continuous BGA mutation schemes (DMS and CMS). models of GAs on this set of functions. With the aim to de ne formally these operators, let The paper is organised as follows. In Section 2 the BGAs are described while in Section 3 a brief review of the parallel x = (x1; : : :; xn) and y = (y1 ; : : :; yn ) be the parent approaches to GAs is reported. In Section 4 an outline of sequences then the o spring z is computed according to the test functions is presented. Section 5 is dedicated to the following recombination and mutation schemes: the analysis and the explanation of the experimental results achieved. The last Section contains the nal remarks and 8 DR : z = x _ y future work. < i i i Recombination : : ELR : zi = xi + (yi ? xi) II. The Breeder Genetic Model EIR : zi = xi + i (yi ? xi ) The BGAs are based on the concepts of evolution of 8 X > CMS : z = x range k?1 2?j species and selection typical of GAs; however they di er < j i i i in the fact that the evolution of the population is `driven' Mutation : > j =0 : DMS : zi = xi range 2?(k ) by the breeding mechanism. This consists in the selection, i at each generation, of the T % best elements within the current population of N elements (T % is called truncation where and i are randomly chosen in the interval rate and its typical values are within the range 10% to ? ; 1 + ], rangei is the mutation range for the i ? th 50%). These elements are let free to mate randomly, allele, j is randomly chosen in the set f0; 1g, is randomly

Abstract | In this paper the Breeder Genetic Algorithms are compared against both serial and parallel Genetic Algorithms by using a wide range of optimisation functions taken from the literature. The aim is to investigate how the change of the tness function in uences their problem{ solving capabilities. The experimental ndings show that the Breeder Genetic Algorithms outperform the Genetic Algorithms considered in terms of number of evaluations and robustness.

chosen in the interval 0:0; 1:0]and k is a precision constant. IV. The test functions employed A wider explanation can be found in 5]. We have taken into account the With these de nitions in mind, we may sketch the problems as in 8]: DeJong's testssame set of optimisation F1{F5 9], Rastrigin's, skeleton of a BGA as follows: Schwefel's and Griewangk's functions 6]. DeJong's F1 function is a unimodal function, F2 is a harder multimodal Procedure Breeder Genetic Algorithm function and F3 is a discontinuous `step{ladder'. F4 begin randomly initialise a population of N individuals involves gaussian noise, so we consider it solved when the best string in the population reaches ?2:5. F5 is while (termination criterion not ful lled) do characterised by the presence of several local minima. evaluate goodness of each individual Rastrigin's (F6), Schwefel's (F7) and Griewangk's (F8) save the best individual in the new population functions are reported in the following: select the best T % individuals for i = 1 to N ? 1 do randomly select two elements among T % X2 x ? A cos(2 x ) x 2 ?5:12; 5:12] F6 : f (x) = n A + recombine them so as to obtain one o spring =1 perform mutation on the o spring X p
n i i i

od end

od update variables for termination

F7 : f (x) = V +
n

F8 : f (x) =

X
i=1

x2 Y xi i 4000 ? cos( pi ) + 1 i=1


i=1 n

?xi sin( jxi j)

xi 2 ?512; 512]

xi 2 ?512; 512]

As attains the function sizes we have used values of 20 for F6 and of 10 for F7 and F8. The value of A in F6 is A variety of parallel Genetic Algorithms has appeared in the literature. Some of the parallel implementations 10.0. As regards F7, the value of V depends on the machine introduce a very strong modi cation in the classical used; in our case we have computed it as 4189.829101. GAs, especially with regard to population structure and V. Performance measurement selection mechanism. Moreover, by means of some of The performance of the BGA is compared with that them, it is possible to simulate biological mechanisms that in the classical (sequential) scheme are not easy to obtained in 8] by means of nine codes of parallel GAs. In take into account. This leads to de ne new (parallel) that paper, four global models, four island models and one implementations of genetic models with di erent structural ne{grained model are considered. Namely, the SGA 2], the Elitist{SGA, the CHC 10] and the Genitor 11] characteristics and ability at solving problems. Current literature on parallel genetic algorithms subdi- algorithms are taken into account both as global and vides these models into three categories: global 3], coarse{ island model. The Cellular{GA 12] is considered as ne{ grained 6] and ne{grained 7]. In the rst class of parallel grained model. Since a global parallel model is, from the GAs, the evaluation of individuals and sometimes the ap- evolution point of view, absolutely identical to a normal plication of genetic operators are explicitly parallelised in sequential GA, in this way we compare our BGA against such a way that the sequential scheme remains unchanged. several variants of both serial and parallel GAs, therefore A more sophisticated idea is used in coarse{grained par- the results of this test will hopefully assess the degree of allel GAs. The population is split into a few subpopu- e ectiveness of the BGA for function optimisation. Our algorithm has been executed on all the test functions lations keeping them relatively isolated from each other. This model of parallelisation introduces a migration pro- for 30 times as in 8]. As regards the parameter set cess by means of which some individuals are exchanged we have used for comparison purposes the same values among subpopulations. Two population genetics models as in 8] for the population size (N = 400) and for the for population structures are used in di erent implemen- maximum number of generations (1000). We know from tations of coarse{grained GAs: the island model and the our experience, however, that this population size is too stepping stone model. The population in the island model large for the BGA to behave at its best, and that for the is partitioned into small subpopulations according to the problems at hand a population ranging between 30 and 100 geographic isolation concept and individuals can migrate is more suited. Before we examine our results, it must be to any other subpopulation. In the stepping stone model, pointed out that the con guration space for the BGA is the population is divided in the same way, but migration is much wider than that of any of the discrete techniques. restricted to neighbouring subpopulations. The third ap- In fact, by using BGAs we can deal with continuous proach in parallelising GAs uses ne{grained parallelism. representation of each real variable, and the precision In the ne{grained parallel GAs the population is parti- for the representation of a variable depends only on the tioned into a large number of very small subpopulations. physical representation of the real variable in the machine The ideal case is to have just one individual for every pro- used, which is, for the hardware utilised, 2:22 10?16. With cessing element available. An exaustive review on parallel reference to the discrete algorithms, instead, the precision is related to the number of bits considered for the encoding. GA approaches can be found in 3] .

III. Parallel Genetic Algorithm models

F1 F2 F3 F4 F5 F6 F7 F8 9 107 1015 1072 1010 1060 1051 1072 GAs 10 BGAs 1049 1032 1083 10481 1035 10333 10186 10186
Table 1. The con guration spaces for the discrete GAs and continuous BGAs.

Algorithm
BGA SGA ESGA pCHC Genitor I{SGA I{ESGA I{pCHC I{Genitor Cellular

3.1 0.3 10.9 30.7 7.4 284 28.9 6.8 83 28.4 6.5 153 17.0 4.1 190 41.3 11.2 417 32.3 7.6 81 33.2 7.4 78 23.2 5.3 112 32.5 8.0 105

Gen StdG Gen StdG Gen StdG Gen StdG Gen StdG

F1

F2

F3

F4

F5

8.6 198 55 139 160 253 40 57 94 94

11.9 16.7 15.3 16.9 8.2 22.0 18.3 18.8 12.3 17.9

1.5 4.2 4.1 3.7 2.1 5.3 5.0 4.4 3.6 4.6

9.3 161 153 223 135 405 375 495 208 397

1.4 40 49 104 67 192 197 239 162 204

2.0 14.6 14.3 16.0 7.9 20.3 13.8 16.3 11.2 15.3

0.0 4.4 4.4 3.9 2.5 6.9 4.7 5.3 3.7 4.3

Table 2. The performance of the BGA against nine GAs on DeJong's test functions.

Table 1 reports the con guration spaces in both cases; it can be seen that the BGA spaces are wider than the others by a number of orders of magnitude ranging from 25 to 409. For the Tables 2 and 4 our stop criterion for BGA is the same as in 8]. In it is related to the precision xed by the bit representation. The value of is equal to 0.01 except for the F3 function which is a step{ladder with step size equal to 1.0, for F4 where = ?2:5 due to the Gaussian noise and for F8, where the results have been obtained with a precision of 10?4. In Table 2 it is reported the average number of generations (Gen) and its standard deviation (StdG) that each algorithm takes to solve the DeJong's functions (the pre x I{ individuates the island models). The table evidences that the average number of generations taken by the BGA is smaller by at least one order of magnitude than that required by discrete parallel GAs, apart from the F3 function. This property holds even for the standard deviation. For the sake of completeness we report in Table 3 the performance of BGA within 1000 generations (the same number as in 8]) when a much tighter convergence criterion is used in terms of the average number of generations, the average nal values obtained (AvgV) and the relative standard deviations (StdV); this leads to values extremely di cult to achieve with discretisation, unless a huge number of bits is employed. The same considerations can be made for the functions F6{ F8. Table 4 contains the performance of the BGA against the parallel GAs in terms of number of runs achieving the best (NR) and of the average value of the nal values obtained (AvgV). From the table the excellent results obtained by the BGA capable of achieving in all cases the global best appear evident. It is also remarkable that

F1 F2 F4 Gen 14.7 275.6 12.7 AvgV 5:0 10?11 5:3 10?11 3:32 StdV 2:6 10?11 2:3 10?11 0:21
Table 3. The performance of the BGA with = 10?10 for F1 and F2 and = ?3:0 for F4.

none of the discrete techniques is able to systematically reach the optimum in all the 30 runs, and this holds for all the problems; especially for F8 the number of successful runs is quite low for all GAs. SGA, in particular, never achieves the optimum for any problem. Table 5 expresses the convergence velocity of our BGA in terms of average number of generations, within the 1000 generations, to overcome the nal solutions of the other implementations. Thus, by using BGAs we can reach better solutions in a lower number of generations. Table 6 reports the average nal values and the relative standard deviations obtained by our BGA let free to evolve for all the 1000 generations. As concerns the evolution process of our BGA in Figure 1 a typical trend is reported for F6, F7 and F8. From the gure it can be observed a quasi{linear slope in the rst phase (until values of about 10?5), in accordance with the theory of BGAs 13].
VI. Conclusions

In this paper we have compared the Breeder Genetic Algorithms against both serial and parallel Genetic Algorithms by using a wide range of optimisationfunctions. Experimental ndings indicate that the former approach outperforms all the other approaches considered for all the

Algorithm
BGA SGA ESGA pCHC Genitor I{SGA I{ESGA I{pCHC I{Genitor Cellular

NR AvgV NR AvgV NR AvgV

F6

F7

F8

30 0 2 23 0 0 13 10 23 24

0 6.8 1.5 0.3 7.9 3.8 0.6 0.9 0.2 0.2

30 0 16 15 20 9 13 28 24 26

0 17.4 17.3 5.9 13.2 6.5 2.6 0.2 0.9 0.7

30 0 1 0 3 7 3 3 6 1

0 0.161 0.107 0.072 0.053 0.050 0.066 0.047 0.035 0.106

test functions in terms of both the number of generations and the robustness. Moreover, since our BGA version is not the best possible in terms of the parameter values used, it is reasonable to expect further improvements in the nal version. The excellent ndings reported in this study lead to test the Breeder Genetic Algorithms on problems with very larger search space. All the results in this paper relative to the discrete GAs have been obtained by V. Scott Gordon and D. Whitley and published in 8]. 1] J. H. Holland, Adaptation in Natural and Arti cial Systems. MIT Press, 1975. 2] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning. Addison{ Wesley, Reading, Mass., 1989. 3] E. Cantu{Paz, \A summary of research on parallel genetic algorithms," Tech. Rep. 95007, IlliGAL Report, July 1995. 4] H. Muhlenbein and D. Schlierkamp{Voosen, \Analysis of selection, mutation and recombination in genetic algorithms," Neural Network World, vol. 3, pp. 907{ 933, 1993. 5] H. Muhlenbein and D. Schlierkamp{Voosen, \Predictive models for the breeder genetic algorithm I. Continuous parameter optimization," Evolutionary Computation, vol. 1, no. 1, pp. 25{49, 1993. 6] H. Muhlenbein, M. Schomisch and J. Born, \The parallel genetic algorithm as function optimizer," Parallel Computing, vol. 17, pp. 619{632, 1991. 7] B. Manderick and P. Spiessens, \Fine{grained parallel genetic algorithms," in Proc. of the 3rd International Conference on Genetic Algorithms, (J. D. Scha er, ed.), pp. 428{433, Morgan{Kaufmann, 1989. 8] V. Scott Gordon and D. Whitley, \Serial and parallel genetic algorithms as function optimizers," in Kaufmann, 1993. 9] K. DeJong, \An analysis of the behavior of a class of genetic adaptive systems," PhD thesis, University of Michigan, 1975. 10] L. Eshelman, \The CHC Adaptive Search Algorithm," 11] D. Whitley and T. Starkweather, \GENITOR II: a distributed genetic algorithm," Journal Expt. Theor. Artif. Intell., vol. 2, pp. 189{214, 1990. 12] V. Scott Gordon, D. Whitley and A. Bohm, \Data ow parallelism in genetic algorithms," Parallel Problem Solving from Nature 2, North Holland, 1992. 13] H. M. Voigt, H. Muhlenbein and D. Cvetkovic, \Fuzzy recombination for the continuous bredeer genetic algorithm," in Proc. of the 6th International Conference on Genetic Algorithms, Morgan{Kaufmann, 1995.
Foundations of Genetic Algorithms and Classi er Systems, Morgan{Kaufmann, 1991. Proc. of the 5th International Conference on Genetic Algorithms, (S. Forrest, ed.), pp. 177{183, Morgan{
References Acknowledgement

Table 4. The performance of the BGA against nine GAs on F6, F7 and F8 with = 10?2.

F6 F7 F8 Gen 191.6 76.6 41.9 AvgV 5:7 10?3 7:6 10?3 7:7 10?5 StdV 2:9 10?3 1:4 10?3 1:4 10?5
Table 5. The performance of the BGA with the same stop criterion as for discrete GAs ( = 10?2 for F6 and F7 and = 10?4 for F8).

F6 F7 F8 Gen 685.4 660.9 555.7 AvgV 4:88 10?6 4:66 10?6 2:79 10?6 StdV 0:91 10?6 1:30 10?6 0:63 10?6
Table 6. The performance of the BGA with = 10?6 .

10000 1000 100 10 1 f(x) 0.1 0.01 0.001 0.0001 1e-05 1e-06 0 100 200 300 400 500 600 Generations 700 800 900 1000 Griewangk Schwefel Rastrigin

Fig. 1. Typical BGA evolutions on the F6, F7 and F8 functions.

You might also like