Choosing the Right Technique for the Right Restriction – A Domain-Specific Approach for Enforcing Search-Space Restrictions in Evolutionary Algorithms
Evolutionary algorithms are a well-known tool for optimising problems that are hard to solve analytically. They mirror the evolutionary approach of recombination and mutation as well as a selection process according to the fitness of an individual. Individuals who violate set search space restrictions are either killed at birth or penalised in their fitness calculation. Which possibility is best to choose depends on the problem at hand and therefore subject to change. Furthermore, restrictions can be vague, for example, when stemming from experiments. We propose a noise-sensitive penalty for violating restrictions and develop a framework where an expert might choose which penalising technique to choose for what kind of restriction. We evaluate our configurable approach against configurations where one technique is used for every type of restriction and find that our approach achieves better results than a strict configuration. Additionally, the noise-sensitive penalising method allows individuals to survive, which may only violate the given restrictions due to a noised testing environment, leading to better results.