Comparing Approaches for Evolving High-Level Robot Control Based on Behaviour Repertoires

摘要

Evolutionary robotics approaches have traditionally been focused on monolithic controllers. Recent studies on the evolution of hierarchical control have, however, yielded promising results. Hierarchical approaches typically rely on a repertoire of behaviour primitives (which themselves can be the result of an evolutionary process), and an evolved top-level arbitrator that continually executes primitives from the repertoire to solve a given task. In this paper, we compare different controller architectures for the evolution of top-level arbitrators. We propose two new methods, one based on neural networks and another based on decision trees induced by genetic programming. We compare the new approaches with existing ones, namely neural network regressors and non-hierarchical control, in a challenging simulated maze navigation task that requires a broad diversity of primitives. Based on empirical results, we draw a number of conclusions regarding the strengths and limitations of each of the studied approaches.