A Dynamic Allocation Method of Basis Functions in Reinforcement Learning