Self-adaptative and Coevolving Memetic Algorithms
Results from applications of meta-heuristics, and Evolutionary Computation in particular, have led to the widespread acknowledgement of two facts. The first is that evolutionary optimisation can be improved by the use of local search methods, creating so-called Memetic Algorithms. The second is that there is no single “best” choice of memetic operators and parameters- rather the situation changes according to both the problem and the particular stage of search. This has created a growing interest in “Adaptive” Memetic Algorithms which combine a portfolio of local search operatorswith some method to choose between them. Here we describe techniques which extend these ideas to allow the behaviours of the local search operators to adapt during the search process. In the first case these maybe thought of as Self-Adaptive, so that each member of the evolving population encodes for both an initial solution to a problem, and a learning mechanism which acts on that solution to improve it. More generally, we show that these can be treated as separate coevolving populations of “genes” and “memes” . Following a review of related work, we next describe a framework formeme-gene self-adaptation and co-evolution. This is followed by a summary of the “proof-of-concept” and of findings concerning representation and scalability with self-adaptive memes. Next the paper considers in more depth issues relevant to co-evolution such as credit assignment, and the ratio of population sizes - which can be thought of as the memetic “load” that an evolving population can support.
KeywordsGenetic Algorithm Local Search Memetic Algorithm Move Operator Steep Ascent
Unable to display preview. Download preview PDF.