NeurIPS 2020

Online Bayesian Goal Inference for Boundedly Rational Planning Agents

Meta Review

This paper consider the task of inferring goals based on agent trajectories, under the assumption that trajectories are generated from an agent with bounded computational power. The authors model the agent using a probabilistic program that interleaves search and execution-by-replanning and perform inference using sequential inverse planning search (SIPS), an SMC method with rejuvenation. Reviewers overall leaned towards acceptance. The main concerns that reviewers raised centered on the fact that the proposed method assumes deterministic dynamics as well as well as knowledge of the planning strategy that is used by the agent. This leads to potential for model misspecification when, e.g., the true agent and the modeled agent employ a different planning algorithm. More broadly reviewers expressed some skepticism about the model of bounded rationality that is adopted in this approach. The authors provide some additional results in their author response that to an extent address concerns about model mismatch. The author response was positively received by reviewers, several of whom raised their scores, whilst providing a caveat that new results were somewhat difficult to evaluate given length constraints of the author response. The AC has carefully read the reviews, authors response, and discussion, and has also done a read-through of the submission. On balance the AC regards this paper as above the bar for acceptance, provided the authors take comments from reviewers regarding model misspecification and difficulty of model estimation to heart, and explicitly acknowledge these as limitations of the work in its current form.