Najmeh Alibabaie - Electrical & Computer Engineering Department, Kharazmi University, Tehran, Iran
Mir Mohsen Pedram - Electrical & Computer Engineering Department, Kharazmi University, Tehran, Iran

This paper deals with a new algorithm for solving the problems of reinforcement learning in continuous spaces. For continuous states and actions, we use a new method based on self-organizing neural network, DIGNET. Two self-organizing neural network, DIGNETs, are used in this method, which having simple structure, can give an appropriate approximate of state/action space. The network is able to adapt inconsistent nature of reinforcement learning environments fine since the system parameters in DIGNET are self-adjusted in an autonomous way during the learning procedure. Automatic and competitive production and elimination of attraction wells, considering parameters of attraction well, threshold, age and depth, lead to flexibility of proposed algorithm to solve continuous problems and finally, the attraction wells of output DIGNET (action) will concentrate on the areas with high reward, providing an appropriate illustration for a continuous state space. At the end, we also present the results of evaluating proposed method on several problems

Machine Learning, Reinforcement Learning, Generalization, Q-learning, DIGNET