Reinforcement learning based on spatial and temporal association of states

In this paper, mechanisms in human learning are incorporated into the reinforcement learning to improve the learning efficiency. A new learning method is presented based on the spatial and temporal association of states, which is inspired by the analogy and recall in human learning. The fuzzy state is proposed to represent the spatial association of states in the state space. The delayed optimization of the control process is proposed for learning with temporally correlated states. In the experiment, the proposed method is applied to a maze problem, which shows that the proposed method has improved learning performance.