Fuzzy wavelet network with reinforcement learning: Application on underactuated system

This paper presents a novel approach of reinforcement learning for continuous systems. The scheme is based in wavelet networks to approximating the continuous space of states. The structure of the wavelet network is dynamically generated accord to the explored regions and trained with a modified Q-Learning algorithm. The wavelet network include a fuzzy inference system which computes the value of the set of possible actions, in order to deal with continuous actions. This novel approach is called adaptive wavelet reinforcement learning control (AWRLC). Simulations of applying the proposed method to underactuated systems are performed to demonstrate the properties of the adaptive wavelet network controller.