Supportive coevolution

Abstract

Automatically configuring and dynamically controlling an Evolutionary Algorithm's (EA's) parameters is a complex task, yet doing so allows EAs to become more powerful and require less problem specific tuning to become effective. Supportive Coevolution is a new form of Evolutionary Algorithm (EA) that uses multiple populations to overcome the limitations of other automatic configuration techniques like self-adaptation, giving it the potential to concurrently evolve all of the parameters and operators in an EA.As a proof of concept experimentation comparing selfadaptation of n uncorrelated mutation step sizes with Supportive Coevolution for mutation step sizes was performed on the Rastrigin and Shifted Rastrigin benchmark functions. Statistical analysis showed Supportive Coevolution outperforming self-adaptation on all but one of the problem instances tested. Furthermore, analysis of instantaneous mutation success rate showed that this new technique is better able to adapt to the changes in the population fitness. Further study using multiple evolving parameters is needed to fully test Supportive Coevolution, but the results presented here show a promising outlook.

Hyper-Heuristics

“…The parameters of these three sets can be seen in Table 2. These parameters were chosen to be consistent with a recent publication using NK-Landscapes [5]. The data used to analyze the scalability of this hyper-heuristic was gathered by running each problem configuration 10 times.…”

Using supportive coevolution to evolve self-configuring crossover

“…SuCo is a method of automatic, self configuration which consists of a primary population, and any number of support populations [4]. The primary population is the population of solution candidates that are evaluated by the target fitness function identically to a traditional EA.…”

“…The support individuals who survive are then allowed to create offspring and continue the evolution of their population in an attempt to continue to find the optimal parameters and operators. As in [4], support individual fitness assignment is based on two factors: the fitness difference between parents and child and the genetic difference between parents and child. Equation 1 defines the fitness difference between parents and their offspring.…”

“…This equation is a simplification for two parent systems of the equation used in [4]. The distance referred to by Equation 2 is defined as:…”

“…SuCo has the potential to evolve more than a single parameter. Since it has been shown that evolving a parameter with SuCo produces promising results [4], this paper describes research aimed at discovering if evolving multiple populations of support individuals by means of SuCo is a feasible approach to further automating EA configuration. Additionally, this research has another novel aspect.…”