A hybrid method for extraction of logical rules from data has been developed. The hybrid method is based on a constrained multi-layer perceptron (C-MLP2LN) neural network for selection of relevant features and extraction of preliminary set of logical rules, followed by a search-based optimization method using global minimization technique. Constraints added to the cost function change the MLP network smoothly into a network performing logical operations. The method is applicable for symbolic and continuos features, finding optimal linguistic variables. Results for several medical and other data sets show that such hybrid technique finds very simple and highly accurate rules, frequently giving results that are more accurate than those obtained by any other classifier. Crisp logical rules are found first, followed by fuzzy rules only if the accuracy of the crisp rules is not satisfactory. Comparison with other rule extraction methods shows superiority of the hybrid approach. The method is also applicable in data mining problems.

Paper in PDF, 95 KB

Projects on similar subject and BACK to the on-line publications of W. Duch.