Evolving a single scalable controller for an octopus arm with a variable number of segments

Abstract

While traditional approaches to machine learning are sensitive to high-dimensional state and action spaces, this paper demonstrates how an indirectly encoded neurocontroller for a simulated octopus arm leverages regularities and domain geometry to capture underlying motion principles and sidestep the superficial trap of dimensionality. In particular, controllers are evolved for arms with 8, 10, 12, 14, and 16 segments in equivalent time. Furthermore, when transferred without further training, solutions evolved on smaller arms retain the fundamental motion model because they simply extend the general kinematic concepts discovered at the original size. Thus this work demonstrates that dimensionality can be a false measure of domain complexity and that indirect encoding makes it possible to shift the focus to the underlying conceptual challenge. © 2010 Springer-Verlag.