The Compact Memetic Algorithm

Abstract

Optimization by probabilistic modeling is a growing research field in evolutionary computation. An example is the compact genetic algorithm (cGA), in which the population of a genetic algorithm (GA) is represented as a probability distribution over the set of solutions. Both cGA algorithm and the order-one behavior of a simple GA with uniform crossover are operationally equivalent. The cGA is much easier to implement and requires less memory. In this paper, memetic algorithms (MAs) are investigated in which the population is replaced by a probability vector analogously to the cGA. The resulting compact memetic algorithms (cMAs) hence require less memory, are easier to implement and require fewer parameters than other MAs. It is shown that cMAs with and without additional recombination perform comparable to or better than population-based MAs on a set of benchmark instances of the unconstrained binary quadratic programming problem.