D. Opitz & J. Shavlik (1994).
Genetically Refining Topologies of Knowledge-Based Neural Networks. International Symposium on Integrating Knowledge and Neural Heuristics, pp. 57-66, Pensacola, FL.

This publication is available in PDF and available in postscript.

Abstract:

Traditional approaches to connectionist theory refinement map the dependencies of a domain-specific rulebase into a neural network, then refine these reformulated rules using neural learning. These approaches have proven to be effective at classifying previously unseen examples; however, most of these approaches suffer in that they are unable to refine the topology of the networks they produce. Thus, when given an ``impoverished'' domain theory, they generalize poorly. A recently published improvement to these approaches, the TopGen algorithm, addressed this limitation by heuristically searching expansions to the knowledge-based networks produced by these algorithms. We show, however, that TopGen's search is too restricted. In response, we present the REGENT algorithm, which uses genetic algorithms to broaden the type of networks seen during its search. It does this by using (a) the domain theory to help create an initial population and (b) crossover and mutation operators specifically designed for knowledge-based networks. Experiments on three real-world domains indicate that our new algorithm is able to significantly increase generalization when compared to both TopGen and a standard approach that does not alter its knowledge-based network's topology.

Return to the publications of the Univ. of Wisconsin Machine Learning Research Group.