A pruning method for accumulated rules by genetic network programming

Genetic Network Programming (GNP) is a newly developed evolutionary computation method. A GNP based rule accumulation method (GNP-RA) is also proposed to generate decision rules and accumulate them into the rule pool, which serves as an experience set for agent control problems. Elite individuals are regarded as evolving rule generators and the extracted rules are viewed as solutions, which is different from the conventional evolutionary computation methods. However, even the best individual could generate some bad rules, thus the interesting rules and uninteresting rules are hard to distinguish. This paper proposed a method to prune the uninteresting rules in the rule pool so that the interesting ones could stand out, which helps to increase the accuracy of decision making. The efficiency and effectiveness of the proposed method is verified by the tile-world problem, which is an excellent benchmark in multi-agent systems.