口语对话系统的POMDP模型及求解

许多口语对话系统已进入实用阶段，但一直没有很好的对话管理模型，把对话管理看做随机优化问题，用马尔科夫决策过程（MDP）来建模是最近出现的方向，但是对话状态的不确定性使MDP不能很好地反映对话模型，提出了一种新的基于部分可观察MDP（POMDP）的口语对话系统模型，用部分可观察特性来处理不确定问题，由于精确求解算法的局限性，考察了许多启发式近似算法在该模型中的话用性，并改进了部分算法，如对于格点近似算法，提出了两种基于模拟点的格点选择方法。

POMDP MODEL AND ITS SOLUTION FOR SPOKEN DIALOGUE SYSTEM

It seems that no excellent model is available for the design of dialogue manager although many spoken dialogue systems have come into practical use in recent years. Using Markov decision process (MDP) is an emerging direction that regards the dialogue strategy selection as a stochastic optimization problem. But the MDP model can't fully reflect the characteristics of a dialogue system because of the uncertainty in the dialogue state. Based on the partially observable MDP (POMDP), a new model for a spoken dialogue system is proposed. It uses the concept of partially observable to handle the uncertainty. Due to the limitation of the exact algorithms, emphais is put on heuristic approximation algorithms and their applicability in the dialogue system POMDP. Two methods for grid point selection are proposed in grid based approximation algorithms.