EVOLUTIONARY CREDIT APPORTIONMENT AND ITS APPLICATION TO TIME-DEPENDENT NEURAL PROCESSING

A new approach to training recurrent neural networks is applied to temporal neural processing problems. Our method combines Darwinian variation and selection with a credit apportionment mechanism for assigning credit to individual neurons within the groups of competing networks. Interconnections between the networks allow the outputs of neurons in one network to be available to the neurons in other networks. The firing behavior of the neurons in a variety of networks is compared with the corresponding neurons in high performing networks for specific input contexts. Payoffs accorded to neurons in one network can thus be shared with neurons in other networks. Only the best neurons over the entire repertoire of networks are allowed to pass their crucial function-determining parameters on to other neurons. The algorithm is demonstrated with connectionist-type units on several temporal processing tasks and compared to genetic algorithms.