An Adaptive and Transferable Dialog Management System for Social Aware Task Execution

Abstract

Efficient and acceptable AI agents need to interact and dialog with their users taking into account not just task efficiency but also social preferences in the interaction. In this work we introduce a model for generating dialog in different scenarios. At the base of the system is a dialog management based on POMDP (Partially Observable Markov Decision Process) models. This task specific component can by itself be used to execute the task. We then introduce another level, and define the interface between the levels, to be able to complement the dialog generated to take into account the social preferences of interaction of each particular user. We show how a particular parameterization of the state allows to learn a personalised policy. We further show that with our formalism a social policy learned for a particular user can then be used in other similar tasks without requiring further learning. We present several simulations showing how we can plan multiple tasks, learn social policies and that those policies can be transferred.