Decision-Theoretic Planning for Intelligent User Interfaces

摘要

The course of interaction between a system and a user cannot in general be predicted with certainty. An interface that is able to anticipate the user's actions several steps ahead can better adapt to the situation at hand, steer the interaction in a promising direction, and protect the user from possible pitfalls. Decision-theoretic planning (DTP) (see, e.g., Boutilier, Dean & Hanks, 1999) provides methods for considering the potential future consequences of available actions. The possible courses of interaction are modeled with a Markov decision process (MDP). Instead of a plan, DTP provides the interface with a policy, which specifies the optimal next step (maximizing the expected utility) for each possible state of the system's interaction with the user. We have studied two problems in detail: (1) the presentation of location-aware navigation recommendations (Bohnenberger et al. 2002) and (2) the presentation of situation-dependent operating instructions for a technical device (Jameson et al. 2001). In both scenarios, the interaction between the user and the interface is modeled with a fully observable Markov decision process (FOMDP). Time is considered as a cost factor of the interaction. Important forms of adaptation include the selection of adequate output modalities and the degree of information detail.