A Mixed Strategy for Evolutionary Programming Based on Local Fitness Landscape

The performance of Evolutionary Programming (EP) is affected by many factors (e.g. mutation operators and selection strategies). Although the conventional approach with Gaussian mutation operator may be efficient, the initial scale of the whole population can be very large. This may lead to the conventional EP taking too long to reach convergence. To combat this problem, EP has been modified in various ways. In particular, modifications of the mutation operator may significantly improve the performance of EP. However, operators are only efficient within certain fitness landscapes. The mixed strategies have therefore been proposed in order to combine the advantages of different operators. The design of a mixed strategy is currently based on the performance of applying individual operators. Little is directly relevant to the information of local fitness landscapes. This paper presents a modified mixed strategy, which automatically adapts to local fitness landscapes, and implements a training procedure to choose an optimal mixed strategy for a given typical fitness landscape. The proposed algorithm is tested on a suite of 23 benchmark functions, demonstrating the advantages of this work in that it is less likely to be stuck in local optima and has a faster and better convergence