Data-Driven Adaptation in Memetic Algorithms

Abstract

As has been empirically demonstrated in the previous chapter, an inattentively configured combination of memetics and a base evolutionary algorithm (EA) can potentially lead to below par optimization performance. The typical issues that must be addressed in the design of such memetic algorithms (MAs), often requiring some degree of domain expertise or extensive human intervention in the tuning of control parameters, include, (i) finding the subset of solutions for which local refinements must be carried out, (ii) determining local search intensity (i.e., the computational budget to be allocated for lifetime learning of individuals in the MA), and (iii) defining the lifetime learning method, i.e., meme, to be used for a particular problem at hand given a catalogue of multiple memes (multi-memes) to choose from. In this chapter, we offer a data-driven alternative to tackling some of these issues.