A dynamical systems analysis of the neural basis of behavior in an artificial autonomous agent
Randall D. Beer, John Christopher Gallagher · 1998
This thesis proposes that one may study the neural basis of behavior by applying standard dynamical systems analysis techniques to artificially evolved biologically-plausible nervous systems. It introduces a specific artificial agent to be controlled, and a number of tasks for that agent to perform. Several hundred continuous time recurrent neural networks of various architectures are evolved to allow the agent to accomplish each task. The best performing networks are subjected to various analyses to determine how they generate appropriate control efforts. The author hopes to use the collected experiments to provide support for the following propositions: (1) That it is possible to artificially evolve neural network control systems, specifically, continuous time recurrent neural networks (CTRNNs), that direct interesting behaviors. (2) That the operation of the artificially evolved CTRNNs is understandable using existing mathematical and experimental techniques. (3) That artificially evolved neural networks can be biologically plausible and may be used as the basis of theories about how natural nervous systems produce similar behaviors. (4) That artificially evolved neural networks are implementable using existing technologies and may be used as the basis of completely artificial control systems.