Explanation by General Rules Extracted From trained Multi-Layer Perceptrons

GR2 is a hybrid knowledge-based system where the knowledge acquired in a trained Multi-layer Perceptron is translated into a symbolic and abstract form called general rules. This is based on both white-box and black-box criteria. The extracted rules can be used for inference on a case-by-case basis, explaining how a decision is made. The extracted rules possess both qualitative and quantitative properties of the domain knowledge, thus enhancing the reasoning capability of the system. The methodology for extracting rules from a trained MLP via two heuristics-the Potential Default Set and the Feature Salient Degree-is outlined and the use of the resulting domain rules in case-by-case explanation is described. A number of examples from synthetic domains is considered and the problem of diagnosing malignancy in breast lesions from observed cytopathological features is presented. Here the case explanations are commented upon by a senior pathologist and favourable agreement is found.