Utilizing bias to evolve recurrent neural networks

Since architectures and weights for recurrent neural networks are difficult to design, evolutionary methods may be applied to search the space of such networks. However, for all but trivial problems, this space is very large. Hence, biases are required that guide the search. Here, we investigate solving a smaller related problem to establish such a bias. Networks are specified by trees containing operators that act on nodes (neurons) and edges (connections). We demonstrate the approach on a signal reproduction task that requires internal state. Performance on a small problem size was improved by solving a smaller problem first. By repeatedly applying the principle, versions of the problem were solved that were not solved by a direct approach