Genetic Network Programming with generalized rule accumulation

Genetic Network Programming(GNP) is a newly developed evolutionary computation method using a directed graph as its gene structure, which is its unique feature. It is competent for dealing with complex problems in dynamic environments and is now being well studied and applied to many real-world problems such as: elevator supervisory control, stock price prediction, traffic volume forecast and data mining, etc. This paper proposes a new method to accumulate evolutionary experiences and guide agent´s actions by extracting and using generalized rules. Each generalized rule is a state-action chain which contains the past information and the current information. These generalized rules are accumulated and updated in the evolutionary period and stored in the rule pool which serves as an experience set for guiding new agent´s actions. We designed a two-stage architecture for the proposed method and applied it to the Tile-world problem, which is an excellent benchmark for multi-agent systems. The simulation results demonstrated the efficiency and effectiveness of the proposed method in terms of both generalization ability and average fitness values and showed that the generalized rule accumulation method is especially remarkable when dealing with non-markov problems.