Adaptive Learning by Emergent Dynamics of Robots
Sadayoshi Mikami · The proceedings of the JSME annual meeting · 2000
Interaction of simple machines often causes chaotic fluctuation. Optimising behaviours for these machines should incorporate how to model the entire system. But it is difficult because the optimization method itself produces the behaviour of the target that should be modeled. To this end, we have proposed a modeling method that uses low-dimensional phase-space embedding and that incorporates reinforcement learning. Although theoretical proof is not fully given, we have applied the method to some of the realistic targets that have stochastic noise. This report shows a case where cellular automata based traffic simulation is used as a test-bed.