پرهام مرادی دولت آبادی

Incorporating skills in reinforcement learning methods results in ac-celerate agents learning performance. The key problem of automatic skill discovery is to find subgoal states and create skills to reach them. Among the proposed algorithms, those based on graph centrality measures have achieved precise results. In this paper we propose a new graph centrality measure for identifying subgoal states that is crucial to develop useful skills. The main ad-vantage of the proposed centrality measure is that this measure considers both local and global information of the agent states to score them that result in iden-tifying real subgoal states. We will show through simulations for three bench-mark tasks, namely, “four-room grid world”, “taxi driver grid world” and “soccer simulation grid world” that a procedure based on the proposed centrality measure performs better than the procedure based on the other centrality measures.