NeurIPS 2020

Meta-trained agents implement Bayes-optimal agents

Meta Review

The reviewers are all positive on this paper. However, I am skeptical about the significance of results on the synthetic datasets considered. I understand that the point is to compare Bayes-optimal performance with RNN performance and hence one must be able to determine Bayes-optimal performance. But these analytically solvable tasks seem fundamentally different from problems of real significance. Hence the acceptance recommendation.