Efficient model learning for dialog management

Abstract

Partially Observable Markov Decision Processes (POMDPs) have succeeded in many planning domains because they can optimally trade between actions that will increase an agent's knowledge about its environment and actions that will increase an agent's reward. However, POMDPs are defined with a large number of parameters which are difficult to specify from domain knowledge, and gathering enough data to specify the parameters a priori may be expensive. This work develops several efficient algorithms for learning the POMDP parameters online and demonstrates them on dialog manager for a robotic wheelchair. In particular, we show how a combination of specialized queries ("meta-actions") can enable us to create a robust dialog manager that avoids the pitfalls in other POMDP-learning approaches. The dialog manager's ability to reason about its uncertainty -- and take advantage of low-risk opportunities to reduce that uncertainty -- leads to more robust policy learning.