Optimizing Dialogue Strategy in Large-Scale Spoken Dialogue System: A Learning Automaton Based Approach

Application of statistical methodology to model dialogue strategy in spoken dialogue system is a growing research area. Reinforcement learning is a promising technique for creating a dialogue management component that accepts semantic of the current dialogue state and seeks to find the best action given those features. In practice, increase in the number of dialogue states, much use of memory and processing is needed and the use of exhaustive search techniques like dynamic programming leads to sub-optimal solution. Hence, this paper investigates an adaptive policy iterative method using learning automata that cover large state-action space by hierarchical organization of automaton to learn optimal dialogue strategy. The proposed approach has clear advantages over baseline reinforcement learning algorithms in terms of faster learning with good exploitation in its update and scalability to larger problems.