Multiple interacting programs: A representation for evolving complex behavior

This paper defines a representation for expressing complex behaviors called multiple interacting programs (MIPs) and describes an evolutionary method for evolving solutions to difficult problems expressed as MIPs structures. The MIPs representation is a generalization of neural network architectures that can model any type of dynamic system. The evolutionary training method described is based on an evolutionary program originally used to evolve the architecture and weights of recurrent neural networks. Example experiments demonstrate the training method's ability to evolve appropriate MIPs solutions for difficult problems. An analysis of the evolved solutions shows their dynamics to be interesting and nontrivial.