A new genetic approach to universal rule generation from trained neural networks
Minoru Fukumi, Yasue Mitsukura, Norio Akamatsu · 2002
A new rule generation method from neural networks is presented. A neural network (NN) is formed using a genetic algorithm (GA) with virus infection and deterministic mutation to represent regularities in training data. This method utilizes a modular structure in GA. Each module learns a different neural network architecture, such as sigmoid and a higher order neural networks. Those chromosome information is communicated to the other modules by the virus infection. The higher order units are connected to an output unit or hidden units. By using these architectures, rules can be extracted. The results of computer simulations show that this approach can generate obvious network architectures and as a result simple rules.