Maintaining Diversity in EDAs for Real-Valued Optimisation Problems

A recent extension applicable to a wide range of discrete EDA algorithms, called sampling-mutation, has shown promise on a non-stationary problem, as well as on a hierarchical deceptive problem. In this paper we further the empirical exploration on Ackley, Rosenbrock and Schwefel, three well-known real-valued variable optimisation problems. The EDA on which we perform our experiments is based on learning and simulation of a Bayesian classifier. The population is at each generation divided into classes based on fitness. The benefit that such classes can have on the diversity of the population and also on the performance of the algorithm, will be evaluated and compared to sampling-mutation. We will show that sampling-mutation can significantly increase the performance of a discrete EDA on said problems by maintaining a higher level of useful population diversity. We also show that an EDA with the use of sampling-mutation can be competitive against a generational genetic algorithm on this type of problem.