Nonstrict Hierarchical Reinforcement Learning for Interactive Systems and Robots

摘要: Conversational systems and robots that use reinforcement learning for policy optimization in large domains often face the problem of limited scalability. This has been addressed either by using function approximation techniques estimate approximate true value a or hierarchical decomposition task into subtasks. We present novel approach dialogue combines benefits both control allows flexible transitions between subtasks to give human users more over dialogue. To this end, each agent hierarchy is extended with subtask transition dynamic state space allow switching subdialogues. In addition, policies are represented linear order generalize decision making situations unseen training. Our proposed evaluated an interactive conversational robot learns play quiz games. Experimental results, simulation real users, provide evidence our can lead (natural) interactions than strict it preferred users.