Abstract: Automatic construction of abstract action is one of the key technologies in hierarchical reinforcement learning. Skill chaining is a typical algorithm for automatically discovery abstract actions in continuous reinforcement learning domains, but the skill chaining algorithm needs to iterate many times and the convergence speed is slow. This paper presents an abstract action tree construction algorithm based on demonstration trajectories (ACADT). By using a change point detection method, ACADT segment each trajectory into a chain. The chains obtained from the multiple trajectories are merged into an abstract action tree. Experimental results show that ACADT can construct an abstract action tree and faster convergence.


Key words: hierarchical reinforcement learning, demonstration trajectories, abstract action, automatic construction, machine learning
