Self adaptation of parameters for MCPC in genetic algorithms

• Main Authors: Susana Cecilia Esquivel, Héctor Ariel Leiva, Raúl Hector Gallard
• Format: Article
• Language: English
• Published: Postgraduate Office, School of Computer Science, Universidad Nacional de La Plata 2000-03-01
• Series: Journal of Computer Science and Technology
• Subjects: genetic algorithm; self-adaptation; crossover; function optimisation
• Online Access: https://journal.info.unlp.edu.ar/JCST/article/view/1012

As a new promising crossover method, multiple crossovers per couple (MCPC) deserves special attention in evolutionary computing field. Allowing multiple crossovers per couple on a selected pair of parents provided an extra benefit in processing time and similar quality of solutions when contrasted against the conventional single crossover per couple approach (SCPC). These results, were confirmed when optimising classic testing functions and harder (non-linear, non-separable) functions. Despite these benefits, due to a reinforcement of selective pressure, MCPC showed in some cases an undesirable premature convergence effect. In order to face this problem, the present paper attempts to control the number of crossovers, and offspring, allowed to the mating pair in a self-adaptive manner. Self-adaptation of parameters is a central feature of evolutionary strategies, another class of algorithms, which simultaneously apply evolutionary principles on the search space of object variables and on strategy parameters. In other words, parameter values are also submitted to the evolutionary process. This approach can be also applied to genetic algorithms. In the case of MCPC, the number of crossovers allowed to a selected couple is a key parameter and consequently self-adaptation is achieved by adding to the chromosome structure -labels- describing the number of crossover allowed to each individual. Labels, which are bit strings, also undergo crossover and mutation and consequently evolve together with the individual. During the stages of the evolution process, it is expected that the algorithm will return the number of crossovers for which the current population exhibits a better behaviour.