Cyclic evolution: a new strategy for improving controllers obtained by layered evolution

Abstract

Complex control tasks may be solved by dividing them into a more specific and more easily handled subtasks hierarchy. Several authors have demonstrated that the incremental layered evolution paradigm allows obtaining controllers capable of solving this type of tasks. In this direction, different solutions combining Incremental Evolution with Evolving Neural Networks have been developed in order to provide an adaptive mechanism minimizing the previous knowledge necessary to obtain a good performance giving place to controllers made up of several networks. This paper is focused on the presentation of a new mechanism, called Cyclic Evolution, which allows improving controllers based on neural networks obtained through layered evolution. Its performance is based on continuing the cyclic improvement of each of the networks making up the controller within the whole domain of the problem. The proposed method of this paper has been used to solve the Keepaway game with successful results compared to other solutions recently proposed. Finally, some conclusions are included together with some future lines of work.