Experiments Assessing Learning of Agent Behavior using Genetic Programming with Multiple Trees

摘要

In this paper, experiments to assess agent behavior learning are conducted to demonstrate the performance of genetic programming (GP) with multiple trees. Using the methods, each has a chromosome representing agent behavior as several trees. We have proposed two variants using the conditional probability and the island model to improve the methods' performance. In GP using the conditional probability, individuals with high fitness values are used to produce conditional probability tables to generate individuals in the next generation. In GP using the island model, the population is divided into two islands of individuals: one island maintains diversity of individuals. The other emphasizes the accuracy of the solution. Moreover, this paper improves methods to seek the optimal number of executions of each tree in an individual. Those methods are applied to a garbage collection problem and a Santa Fe Trail problem. They are compared with traditional GP, GP with control nodes, and genetic network programming (GNP) with control nodes. Experimental results show that our methods are effective for improving the fitness.