Symbolic Interpretation of Arti cial Neural Networks

H ybrid Intelligent Systems that combine knowledge based and artiicial neural network systems typically have four phases involving domain knowledge representation, mapping of this knowledge into an initial connectionist architecture, network training, and rule extraction respectively. The nal phase is important because it can provide a trained connectionist architecture with explanation power and validate its output decisions. Moreover, it can be used to reene and maintain the initial knowledge acquired from domain experts. In this paper, we introduce three new rule extraction techniques. The rst technique extracts a set of binary rules from any neural network regardless of its kind (MLP, RBF etc.,). The second technique extracts partial rules that represent the most important embedded knowledge in a trained network. The delity of the second technique is adjustable to the desired level of knowledge extraction. The third technique is a comprehensive and universal approach. A rule evaluation technique that orders extracted rules based on three performance measures is then proposed. The three techniques are applied to the iris and breast cancer data sets. The extracted rules are evaluated qualitatively and quantitatively, and compared with those obtained by other approaches.