Utilizing Domain Knowledge in Neuroevolution

James Fan, Raymond Lau, Risto Miikkulainen

Abstract: We propose a method called Rule-based ESP (RESP) for utilizing prior knowledge evolving Artificial Neural Networks (ANNs). First, KBANN-likete chniques are used to transform a set of rules into an ANN, then the ANN is trained using the Enforced Subpopulations (ESP) neuroevolution method. Empirical results in the Prey Capture domain show that RESP can reach higher level of performance than ESP. The results also suggest that incremental learning is not necessary with RESP, and it is often easier to design a set of rules than an incremental evolution scheme. In addition, an experiment with some of the rules deleted suggests that RESP is robust even with an incomplete knowledge base. RESP therefore provides a robust methodology for scaling up neuroevolution to harder tasks by utilizing existing knowledge about the domain.