NeurIPS 2020

Optimally Deceiving a Learning Leader in Stackelberg Games

Meta Review

Four knowledgeable referees agree that the paper should be accepted (though one referee, R3, remains borderline). R3 continues to argue that the amount of time required to learn to deceive the learner can be exponential. While the authors argue that sometimes it would not be, the reviewer's criticism is valid, though this does not invalidate the contribution the paper makes. Thus, I concur that the paper should be accepted. The paper is theoretical in nature, expanding on a wide range of results. I believe the paper could be presented effectively as a poster.