Supporting Multiple User Types with a Multimodal Dialog Agent

Recent research has addressed the problem of formulating a dialog agent as a partially observable Markov decision process (POMDP), and learning a dialog policy that is optimal given the particular characteristics of the transition, observation and reward functions of the POMDP. This paper addresses the problem of trying to learn a small set of dialog agent policies that provide near-optimal behavior over a wide range of variations in POMDPs, reflecting different user preferences and environment characteristics. We show for a very simple dialog, we can cover a large number of simulated users to within 10% of their optimal return using fewer than 5% of the individual optimal policies.