Variations in evolution of subsumption architectures using genetic programming: the wall following robot revisited

Abstract

The wall following robot is examined as a potential benchmark problem for applications of genetic programming (GP) to emergent robotic behavior. This paper describes experiments that were performed to characterize the performance, solution space, search space, and robustness using GP with and without automatically defined functions (ADFs). GP with ADFs was unable to significantly outperform GP without ADFs on this problem. A suboptimal modality was discovered across all four program architectures. Many of the optimal solutions that were discovered tended to limit the number of sensors used for wall-following behavior; some used as few as three sensors. Tests for robustness indicate a "handedness" to the evolved solutions, which does seem to contribute to solution brittleness.