Reening Symbolic Knowledge Using Neural Networks

To use artiicial neural networks as part of a multistrategy learning system, there must be a way for neural networks to accept information, often expressed symbolically, from other systems. Moreover, neural networks must be able to transfer the results of their learning. A three-step process for learning using rule-based domain knowledge in combination with neural networks can address this task. Methods for performing the rst two steps have been previously described. Hence, this chapter focuses on the nal step, extracting symbolic rules from trained neural networks. Proposed and empirically evaluated in this chapter is a new method for rule extraction. The rules extracted by this method closely reproduce the accuracy of the network from which they came, are superior to the rules derived by a learning system that directly reenes symbolic rules, and are human comprehensible. Hence this process allows the use of neural networks as a part of a multistrategy learning system. More generally, this work contributes to the understanding of how symbolic and connectionist approaches to artiicial intelligence can be prootably integrated.