On the potential benefits of knowing everything

Abstract

Evolutionary algorithms are designed to find impressive solutions in complex search spaces. Meeting this aim requires that the heuristic guiding search aligns with the structure of the search space, i.e. the effectiveness of rewarding properties of individuals (like fitness or novelty) depends on how those properties are distributed. Interestingly, researchers can rarely access ground truth about such connectivity, especially in settings like evolutionary robotics (ER) where search spaces are large and an individual's behavior could potentially inform search in many different ways. This paper raises the intriguing possibility of adapting or simplifying existing ER domains such that we know everything about the search space's structure, to enable us to develop intuitions and quickly explore new search algorithms. The proposed approach is to pair an expressive (but limited) encoding with a benchmark ER domain, and precompute the behavior of all possible individuals. Such precomputation enables evaluation as a look-up table, and the further precomputation of normally-intractable quantities, like exact rarity of behaviors and a variety of evolvability metrics. Evolution can then be driven and gauged by such properties with extreme efficiency. The hope is that insights gleaned from this sandbox can inspire new and effective approaches that generalize to when everything is not known.