Abstract
Recently, a lot of methods of using the neural network for the state space construction of a mobile robot are proposed. When robot is put on a different environment, it is not possible to behave robustly because these methods make a robot adjust to one static environment. On the other hand, Subsumption Architecture (SA) is not suitable for tasks of depending on the structure of an environment though it is expected that it can behave robustly even by the dynamic environment. The robustness of SA is declined when robot adjusts to a specific environment by neural network. In this paper, we propose the hybrid model which is consisted of SA and Dynamics-Based Self-organizing Incremental Neural Network (DBSOINN). DBSOINN is modified The Self-organizing Incremental Neural Network (SOINN) for state space construction of the reinforcement learning. The effectiveness of this proposal was confirmed by the simulation experiment that the mobile agent behaves in the environment which is composed of plural mazes. The proposed model is able to use plural DBSOINN appropriately at the maze which changes dynamically.
