Hierarchical Dialogue Policy Learning using Flexible State Transitions and Linear Function Approximation

摘要

Conversational agents that use reinforcement learning for policy optimization in large domains often face the problem of limited scalability. This problem can be addressed either by using function approximation techniques that estimate an approximate true value function, or by using a hierarchical decomposition of a learning task into subtasks. In this paper, we present a novel approach for dialogue policy optimization that combines the benefits of hierarchical control with function approximation. The approach incorporates two concepts to allow flexible switching between subdialogues, extending current hierarchical reinforcement learning methods. First, hierarchical treebased state representations initially represent a compact portion of the possible state space and are then dynamically extended in real time. Second, we allow state transitions across sub-dialogues to allow non-strict hierarchical control. Our approach is integrated, and tested with real users, in a robot dialogue system that learns to play Quiz games.