A topology exploiting genetic algorithm to control dynamic systems

Abstract

Learning to control systems with unknown dynamics is a challenging task for machine learning algorithms. One of the problems that arises here is that the input space is a high dimensional real vector space, while many learning algorithms usually work with a preferably small set of discrete units. Partitioning the input space into a small proper set of subspaces however requires some domain knowledge which is not available to the learning algorithm, while choosing a relatively large set of subspaces makes learning a lot more difficult. In this paper we examine the feasibility for genetic algorithms to exploite the continuity of the input space, such that learning is made more efficient for larger sets of subspaces. Every gene of the genotype represents one subspace, while the alleles specify the action that has to be taken in that part of the input space. The genes are structured on the genotype such that adjacent genes are representatives of neighboring points in the higher dimensional input space. By exploiting this property the GA is capable of finding a solution quite efficiently although the genotype has considerable length.