Coevolution as an Autonomous Learning Strategy for Neuromodules

The training of artificial neural networks to solve a particular problem usually affects only the weight space of the network. In contrast, the network architecture has to be predefined by the user. For this pretraining choice of network architecture profound skill and experience may be required on the part of the user — an improperly predefined architecture easily renders a learning problem insoluble. Relevant parameters of a network architecture include the choice of feedforward or recurrent structure, of network size and connectivity level, and of the number of layers. The appropriateness of such choices can only be determined posttraining. Consequently, methods have been devised either for determining upper bounds for network sizes (e.g. the application of Kolmogorov’s theorem to 3-layered network structures by Hecht-Nielsen) or for automatically pruning neurons from an initial structure in order to adjust network size (e.g. Bartlett’s dynamic node architecture learning). However, these methods are restricted to a particular subclass of network architectures involving only feedforward connections. Here we suggest a general approach allowing for a joint adaptation of neural networks with respect both to weight dynamics and architecture dynamics. All the relevant parameters of network architecture mentioned above are subject to the learning process and thus are adjusted to the specific problem considered. This applies in particular to the crucial alternative of feedforward or recurrent network structure, in fact, the proper choice between these paradigms is automatically made.