Practical dialogue manager development using POMDPs

Partially Observable Markov Decision Pro- cesses (POMDPs) are attractive for dialogue management because they are made to deal with noise and partial information. This pa- per addresses the problem of using them in a practical development cycle. We apply fac- tored POMDP models to three applications. We examine our experiences with respect to design choices and issues, and compare per- formance with hand-crafted policies.