GSO-2010: Steven Williams and Larry Yaeger
Dynamical Systems Analysis of Evolving Neural Networks
From man-made computational systems to biological brains, low levels of chaos have been shown to aid information processing. This talk examines the evolved neural networks of Polyworld, an artificial life simulation, from a dynamical systems perspective. An analysis of bifurcation diagrams and Lyapunov exponents shows that evolution drives these networks toward a dynamical phase transition between order and disorder. This trend is actively selected for by evolution while the agents behaviorally adapt to their environment; comparison against a random null model indicates that the trend is statistically significant. Routes to chaos are varied, but similarities in the bifurcation diagrams hint at the possibility of equivalence classes of dynamical neural behavior. Possible future work will examine this possibility and investigate relationships between Lyapunov exponents and relevant information- and graph-theoretic measures.