Abstract
This paper presents a method to extract rules from multilayered neural networks (NNs) formed using evolutionary algorithms (EAs). The objective of this study is to extract rules from NN, achieving 100% recognition accuracy in pattern recognition systems. NNs to be extracted rules are formed using EAs. Hy-brid algorithms of NN and EAs perform the formation of small-sized NN systems, which are suitable for rule extraction. EAs in this paper are a random optimization (search) method (ROM) and a genetic algo-rithm (GA). In this paper iris data and coin data sets are used as inputs. EAs are utilized to reduce the number of connection weights in NNs. The network weights survived after the training by ROM and GA represent regularities to perform pattern classification. The rules are then extracted from the networks in which hidden units use signum functions to produce binary hidden outputs, while sigmoid units are used in NN training. It enables us to extract simple logical functions from the networks. It is shown by means of computer simulations that this approach is more effective in rule extraction than conventional methods.
