Feedback-Guided Intention Scheduling for BDI Agents

Intelligent agents, like those based on the popular BDI agent paradigm, typically pursue multiple goals in parallel. An intention scheduler is required to reason about the possible interactions between the agent's intentions to maximize some utility. An important consideration when scheduling intentions is the user's preferences over the goals and the ways in which the goals are achieved. These preferences are generally unknown in advance, time-consuming to elicit, hard to model, and difficult to incorporate into an intention scheduler. In this paper, we present a Monte Carlo Tree Search based intention scheduler (pref-MCTS) that is able to learn the user's preferences over intention schedules via low-burden comparative-type queries. It incorporates the learned preferences in guiding the search, leading to execution policies that are optimized towards the user's preferences and expectations. We evaluate our approach using an artificial oracle that shows that pref-MCTS improves over state-of-the-art baselines, even when provided with a limited number of preference queries and noisy labels. We also conducted a user study and showed that pref-MCTS is able to learn user preferences and generate schedules that are preferred by the users in real-time.