Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
This work is original, to the best of my knowledge, in automatically grading an agents capability in a Goal directed MDP. This allows the agent to generate increasingly difficult goals to perform by solving a Wasserstein barycenter problem; they effectively explain this problem as solve it well by exploiting some structure in the MDP. The authors also provide a simple, intuitive setting where the effects of HGG are conveyed, but do not include much discussion of their more complex experimental results. While I do not doubt the validity of their approach, when experimental results are included it seems odd to not try and dissect what their method was able to contribute. Their ablation studies (in the main body as well as the supplemental material) offer additional evidence that their approach is indeed valid; one should expect that their automatic method should be robust to a reasonable range of hyper-parameters, which was important to demonstrate.
update: I've read the rebuttal and keep my original score. It is good to see new results on a harder environment + how choosing right metric helps in another one. I keep the original score since the theory/math is sufficient for this method, but is not a significant contribution. --- The authors propose a novel curriculum generation technique that uses both the observed task goals/initial states and the visited states information to automatically generate a set of achievable but task-relevant goals for better exploration. The technique is combined with HER to achieve substantial improvements on sparse reward domains, especially in the context with adversarial initial-goal states. The paper is written very cleanly and in good quality. The motivation is well grounded based on the smoothness of value functions. The idea is simple but novel and effective. The illustration in Figure 2 is convincing of the failure of HER in adversarial cases and how this approach can improve exploration. Figure 4’s comparison with prior ground-truth success-based curriculum also shows favorable result. This curriculum method is also realistic, as it does not make resettability assumption or Monte Carlo success rate evaluation as done in other works such as reverse curriculum generation. The author also discusses the core limitation of the method as the reliance on distance metric for curriculum and suggests future work in conclusion. Comments: - extension: for faster adaptation to target goal distribution, authors may even consider changing exploration goal within a rollout (in K=1 setting) - can monotonic improvement guarantee for this curriculum approach be derived?
The authors propose a new method for sampling exploration goals when performing goal-conditioned RL with hindsight experience replay. The authors propose a lower bound that depends on some Lipschitz property of the goal-conditioned value function with respect to the distance between the goals and states. The authors demonstrate that across various Fetch-robot tasks, their method, when combined with EBP (a method for relabeling goals), outperforms HER. The authors also perform various ablations that show their method is relatively insensitive to hyperparameter values. Overall, the empirical results are solid, but the math behind the paper is rather troubling. Specifically, the theorem seem rather vacuous: Writing “x” in place of “(s, g)”, the theorem basically says that if V(x1) >= V(x2) + d(x1, x2), then if you take the expectation of both sides (w.r.t. any coupling over x1 and x2), the inequality still holds. Taking the minimum overall couplings gives the theorem. Reading the (brief) proof only makes me more confident that the theorem is not insightful. Math aside, the intuition behind the method is clear (especially with Figure 2). The sensitivity analysis and comparisons seem appropriate, though not particularly amazing. Some smaller comments: The first paragraph of the introduction seems rather generic. It would be good to quickly focus on the problems the authors are trying to solve (goal-directed exploration) more quickly. I do wonder how important it is to solve the bipartite matching problem exactly. For example, could the authors have instead sampled 100 different trajectories and taken the max over all time steps and trajectories? The related works could discuss more work such as [1-7], though some of the work is extremely recent and I do not penalize the authors for not including them.  S. Forestier, Y. Mollard, and P.-Y. Oudeyer. Intrinsically Motivated Goal Exploration Processes with Automatic Curriculum Learning.  A. Péré, S. Forestier, O. Sigaud, and P.-Y. Oudeyer. Unsupervised Learning of Goal Spaces for Intrinsically Motivated Goal Exploration.  C. Colas, et. al. GEP-PG: Decoupling Exploration and Exploitation in Deep Reinforcement Learning Algorithms  V. Pong, et. al. Skew-Fit: State-Covering Self-Supervised Reinforcement Learning  V. Veeriah. J Oh., and S. Singh. Many-Goals Reinforcement Learning.  R. Zhao, et. al. Maximum Entropy-Regularized Multi-Goal Reinforcement Learning  Kaelbling, Leslie P. Learning to achieve goals. -- After reading the author feedback, I still find the mathematical results rather weak. Again, the theorem simply says "if the triangle inequality is true for all states, then it is true in expectation, no matter what the (coupling) distribution is. There's no clear explanation for why this assumption would be reasonable, and the only justification provided is that prior related work made this assumption implicitly. Personally, I believe that that it is a reasonable assumption in many applications, and the paper would be strengthened by explicitly discussing this, rather than leaving it to the reader. I'm increasing my score in spite of the distracting mathematical discussion, primarily due to the strong empirical results (performance and visualizations).