A new rule extraction method from neural networks
Minoru Fukumi, Norio Akamatsu · 2003
This paper presents a method of extracting rules from multilayered neural networks (NN) formed using a random optimization (search) method (ROM). The objective of this study is to extract rules from NN, achieving 100% recognition accuracy in a pattern recognition system. NNs to be extracted rules are formed using ROM. A hybrid algorithm of NN and ROM performs a formation of a small-sized NN system, which is suitable for a rule extraction. In this paper iris data is used as inputs. ROM is utilized to reduce the number of connection weights in NN. The network weights survived after the ROM training represent regularities to perform pattern classification. The rules are then extracted from the networks in which hidden units use signum and sigmoid functions to produce binary outputs. It enables us to extract simple logical functions from the network. By means of computer simulation, the effectiveness of this approach is examined.