Rule accumulation method with modified fitness function based on Genetic Network Programming

Genetic Network Programming (GNP) extended from GA and GP is competent for the complex problems in dynamic environments because of its directed graph structure, reusability of nodes and implicit memory function. In this paper, a new method to extract and accumulate rules from GNP is proposed. The general idea is to update the fitness values of the rules accumulatively, rather than just replacing them in the former research. That is, the rules which appear frequently in different generations are given higher fitness values because they represent good universal experiences from the past behaviors. By extracting the rules during the evolutionary period and then matching them with agents´ environments, we could guide the agents properly and get better rewards. In order to test the efficiency and effectiveness of the proposed method, we applied the proposed method to the problem of Tile-world as the simulation environment. Simulation results demonstrate the effectiveness of the proposed method.