Learning of neural networks with parallel hybrid GA using a royal road function

In the learning of neural networks, the hybrid genetic algorithm (GA) is one of useful methods, since it can find an optimal set of weights in shorter timer. However, the GA part requires many individuals in a population to maintain its diversity and then it remains a trade-off between the population size and time. We introduce a new idea of evaluation of its chromosome based on the building block hypothesis. We assume an index with same length of an individual and measure the length of corresponding bits to it. Then, we make a reproduction using both fitness and its new index. Furthermore, we change its length from dynamically short to long according to the convergence situation, since intermediate order schemata results from combination of the lower order schemata. To verify the effectiveness of the proposed method, we developed a medical diagnosis system. It is shown that an optimal solution was found in the population size of 10