Performance of genetic network programming for learning agents on perceptual aliasing problem

In this paper, we examine the performance of genetic network programming (GNP) for learning agents on perceptual aliasing problems. Perceptual aliasing problems (PAP) are known as the problem where a learning agent can not distinguish between differing states of the world due to the limitation of its sensors. In order to cope with this problem, a genetic programming approach called Adaptive Genetic-Programming Automata (AGPA) has been proposed. While it effectively tackled to PAP, too many rules are generated that are not used to control the agent due to its tree-based structure. Using GNP, we can reduce the number of rules for PAP since it has network architecture but tree architecture as used in adaptive GP automata. We compare the performance of GNP and AGPA on a maze problem in which a learning agent tries to reach a goal. Simulation results clearly show that the number of rules can be reduced by GNP.