Evolution and Analysis of Minimally Cognitive Model Agents

Randall D. Beer
Department of Electrical Engineering and Computer Science and Department of Biology
Case Western Reserve University

This is the abstract of a talk prepared for the International interdisciplinary seminar on new robotics, evolution and embodied cognition (IISREEC).12th to 15th November 2002, Lisbon, Portugal

Abstract: It is becoming increasingly clear that situated, embodied and dynamical ideas have an important, perhaps fundamental, role to play in the conceptual structure of cognitive science. Before this role can be fully realized, however, at least two challenges must be met.

The first challenge is to develop a proper theoretical framework for these ideas. In particular, we must construct the concepts and tools necessary to understand the often tangled, counterintuitive brain-body-environment organizations that evolution produces. While good progress is being made on the empirical analysis of particular biological brain-body-environment systems, there is no experimental system in which it is currently feasible to simultaneously study all three of these components. By analogy with the role of such idealizations as frictionless planes in physics, I will argue that achieving such an understanding will require a systematic study of frictionless *brains*: simpler idealized models of nervous systems, bodies and environments interacting to produce coordinated behavior. Toward this end, we have been using evolutionary algorithms to evolve "nervous systems" for model agents and then analyzing their brain-body-environment dynamics using the tools of dynamical systems theory.

The second challenge is to extend this approach to address behavioral tasks that are sufficiently cognitively-interesting to engage cognitive scientists. While somewhat subjective, "cognitive interest" is typically strongly correlated with the sophistication of the behavior in which the agent engages and the degree of disconnection of the agent's behavior from its immediate sensory environment. For this reason, we have been exploring the evolution of model agents exhibiting at least minimally cognitive behavior. I will provide an overview of a series of experiments on the evolution of a variety of visually-guided behavior, including object discrimination, short-term memory and selective attention. I will also briefly describe the kinds of analyses that can be performed on these evolved agents, ranging from behavioral and pyschophysical studies to detailed studies of the evolved neural dynamics. Such analyses illustrate some of the implications and challenges of taking a situated, embodied and dynamical perspective on cognitive behavior.

Related Links

For more information contact Luis Rocha at rocha@lanl.gov
Last Modified: October 31, 2002