A decision-theoretic model of assistance

Abstract

There has been a growing interest in intelligent assis-tants for a variety of applications from organizing tasks for knowledge workers to helping people with demen-tia. In this paper, we present and evaluate a decision-theoretic framework that captures the general notion of intelligent assistance. The objective is to observe a goal-directed agent and to select assistive actions in order to minimize the overall cost. We formulate the problem as an assistant POMDP where the hidden state corre-sponds to the agent’s unobserved goals. This formu-lation allows us to exploit (partial) domain models for both estimating the agent’s goals and selecting assistive action. In addition, the formulation naturally handles uncertainty, varying action costs, and customization to specific agents via learning. We argue that in many do-mains myopic heuristics will be adequate for selecting actions in the assistant POMDP and present two such heuristics—one based on MDP planning, and another based on policy rollout. We evaluate our approach in two domains where human subjects perform tasks in game-like computer environments. The results show that the assistant substantially reduces user effort with only a modest computational effort.