Research on fast reinforcement learning

Based on eligibility trace theory, a delayed fast reinforcement learning algorithm DFSARSA (λ) is proposed. By redefining the eligibility trace and tracking the TD (λ) error, the Q-value of reinforcement learning updates may be postponed when they are needed instead of update in each step as traditional SARSA (λ). The update computing complexity is reduced from O (|S| |A|) to O (|A|) compared with SARSA (λ) and the speed of the reinforcement learning is improved greatly. Simulation results show the method is validity.