Exploring GenNet behaviors-using genetic programming to explore qualitatively new behaviors in recurrent neural networks

Until the recurrent backdrop algorithms came along, there was a widespread belief that no generally acceptable procedure existed to train nonconvergent networks. It is shown that recurrent backdrop is not the only algorithm capable of doing this. The alternative proposed uses the technique of genetic programming (GP), i.e., using genetic algorithms (GAs) to evolve output behavior in neural networks (called GenNets). At least one example of a GenNet is presented for each of three cases of time-dependent/independent inputs/outputs, and it is shown how GP techniques were used to evolve GenNets whose operating conditions satisfied the three cases. Some of the extraordinary properties of time-independent GenNets are discussed. The sophisticated behaviors generated by GenNets and recurrent backdrop algorithms are compared. It is claimed that the GenNet behavior is more flexible and interesting because it does not require the training process to be closely supervised