Extracting DNF rules from artificial neural networks

Abstract

Artificial neural networks are powerful classification mechanisms. Neural networks encode knowledge in a set of numerical weights and biases. This data driven aspect of neural networks allows easy adjustments when change of environments or events occur. Numeric weights, however, are difficult to interpret in terms of rules, making it difficult for a human to understand what the neural network has learned. One approach to understanding the representations formed by neural networks is to extract symbolic rules from networks, since concepts represented by symbolic learning algorithms are more easily understood by humans. It has been shown that most concepts described by humans usually can be expressed as production rules in disjunctive normal form (DNF) notation. Rules expressed in this notation are therefore highly comprehensible and intuitive. A method that extracts production rules in DNF is presented. The extracted rules are accurate and results compare favourably with traditional symbolic rule extraction methods. Since the rules are in a logically manipulatableform, significant simplifications in the structure thereof can be obtained, yielding a highly comprehensible set of rules.