Self-Adaptation in Evolutionary Algorithms for Combinatorial Optimisation

It is well known that the choice of parameter settings for meta-heuristic algorithms has a dramatic impact on their search performance and this has lead to considerable interest in various mechanisms that in some way attempt to automatically adjust the algorithm’s parameters for a given problem. Of course this raises the spectre of unsuitable parameters arising from a poor choice of learning/adaptation technique. Within the field of Evolutionary Algorithms, many approaches have been tried, most notably that of “Self-Adaptation”, whereby the heuristic’s parameters are encoded alongside the candidate solution, and acted on by the same forces of evolution. Many successful applications have been reported, particularly in the sub-field of Evolution Strategies for problems in the continuous domain. In this chapter we examine the motivation and features necessary for successful self-adaptive learning to occur. Since a number of works have dealt with the continuous domain, this chapter focusses particularly on its aspects that arise when it is applied to combinatorial problems. We describe how self-adaptation may be use to control not only the parameters defining crossover and mutation, but also how it may be used to control the very definition of local search operators used within hybrid evolutionary algorithms (so-called memetic algorithms). On this basis we end by drawing some conclusions and suggestions about how this phenomenon might be translated to work within other search metaheuristics.