摘要：Divide and rule is an effective strategy to solve large and complex problems. We propose an approach to make agent can discover autonomously subgoals for task decomposition to accelerate reinforcement learning. We remove the state loops in the state trajectories to get the shortest distance of every state from the goal state, then these states in acyclic state trajectories are arranged in different layers according to the shortest distance of them from the goal state. So, to reach these state layers with different distance to the goal state can be used as the subgoals for agent reaching the goal state eventually. Compared with others, autonomy and robustness are the major advantages of our approach. The experiments on Grid-World problem show the applicability, effectiveness and robustness of our approach.