Tasks Decomposition for Improvement of Genetic Network Programming

Genetic Network Programming is an evolutionary algorithm which can be considered as an extension of Genetic Programming but with a graph-structure instead of tree-structure individuals. This algorithm is mainly used for single/multi-agent decision making. It uses a graph to model a strategy that an agent follows to achieve its goal. However, in this algorithm, crossover and mutation operators repeatedly destroy the structures of individuals and make new ones. Although this can lead to better structures, it may also break suitable structures in elite individuals and increase the time needed to achieve optimal solutions. In this research, we modified the evolution process of Genetic Network Programming so that breaking useful structures will be less likely. In the proposed algorithm, the experiences of the best individuals in successive generations are saved. Then, in some specific generations, these experiences are used to generate offspring. The experimental results of the proposed method were tested on two common agent control problem benchmarks namely Tile-world and Pursuit-domain. The results showed the superiority of our method with respect to standard Genetic Network Programming and some of its versions.