Högskolan i Skövde

This paper discusses the use of recurrent neural networks for control of and learning in robots and autonomous agents. In particular the use of feedback in both first- and higher-order recurrent network architectures for the realization of adaptive robot behavior is investigated. Two experiments, in which controller network weights are evolved to solve tasks requiring robots to exhibit context- or state-dependent behavior, are used to demonstrate and analyze different recurrent control architectures.