Pruning of Redundant Information to Improve Performance for Agent Control in A Changing Environment

抄録

Genetic Network Programming(GNP) is a new evolutionary computation method which is competent for many agent control problems. However, some redundant nodes exist in the program of GNP, which can easily cause the over-fitting problem and decrease its performance. In order to prune these nodes, a new method named “Credit GNP” is proposed in this paper. The novelties are, firstly, Credit GNP has a unique structure, where each node has an additional “credit branch” which can skip the redundant nodes. Secondly, Credit GNP combines evolution and reinforcement learning, i.e., off-line evolution and on-line learning to prune the redundant nodes. Which node to prune and how many nodes to prune are determined automatically considering different environments. Simulation results on the Tile-world problem show that Credit GNP could generate not only better programs, but also more general rules for agent control. The superiority of the proposed method over the conventional GNP, GP and standard reinforcement learning is proved.