Abstract
In general, the structure of neural network, such as the layered type or the mutually connected type, etc., is chosen as to a given problem. And connection weights and thresholds of neural units are usually determined by the learning method corresponding to the structure. Several given problems, however, cannot be always solved by the neural network of the structure decided beforehand. Therefore, Flexibly Connected Neural Network (FCN) was previously proposed as a method of constructing arbitrary neural networks with optimized structures and parameters to solve unknown problems. FCN was applied to action control of an autonomous agent and showed experimentally that it is effective for perceptual aliasing problems. The number of hidden units, however, was determined experimentally in the FCN. In this paper, we propose a method based on FCN, which can determine automatically the number of hidden units without trial and error. In order to verify the effectiveness, we applied the proposed method to Tartarus Problem that requires action control of an autonomous agent in unknown maps and we analyzed the obtained action rules.
