Interactive learning of consensus sequences in genetic programming for evolution of snake-like robot

Abstract

We discuss an approach of incorporating interactively learned consensus sequences (ILCS) in genetic programming (GP) for efficient evolution of simulated Snakebot situated in a challenging environment. ILCS introduce a biased mutation in GP via probabilistic context sensitive grammar, in which the probabilities of applying the production rules with multiple right-hand side alternatives depend on the grammatical context. The distribution of these probabilities is learned interactively from the syntax of the Snakebots, exhibiting behavioral traits that according to the human observer are relevant for the emergence of ability to overcome obstacles. Because at the earlier stages of evolution these behavioral traits are not necessarily pertinent to the best performing (i.e. fastest) Snakebots, the user feedback provides the evolution with an additional insight about the promising areas in the fitness landscape. Empirical results verify that employing ILCS improves the efficiency of GP in that the evolved Snakebots are faster than those obtained via canonical GP.