The paper describes new offline and online techniques to optimize the policy of continuous time discrete state and action POMDPs. This paper makes an important contribution to the RL and control literature. Very little work has focused on continuous time control problems in the ML community. While the techniques assume that the model is known, do not scale to high dimensional problems and were tested only on toy problems, they introduce new formalisms that will help the community get familiar with the mathematics of continuous time control. Hence this paper will be of high interest for the RL community.

This appears to be a detailed, carefully written, and theoretically/empirically supported piece of work. The results are based on RKHS theory and appear to extend the state-of-the-art in a number of ways, including learning KMEs while dealing with dependent samples (e.g. from a Markov chain), low dimensional/rank approximations with state-of-the-art error bounds, and a variational approach for clustering, as well as a robust set of experiments. On the whole, I am satisfied that this clearly meets the acceptance criteria for NeurIPS and I am recommending it for acceptance. I have a few comments below on aspects of the paper, where things were either unclear, or might benefit from a revisit. It is unclear how strong assumption 2 is.

The authors generalize 1-lag coupling of the chains to L-lag coupling and provide upper bounds on some distribution distances including the total variation and 1-Wasserstein distance. This bound serves as a convergence check for MCMC, e.g., to stop the burn-in phase. The main contributions of the paper are 1) deriving a computable bound of the distribution distance between two (L-lagged) chains, and 2) presenting algorithms (e.g., Coupled Random-Walk Metropolis-Hastings, Coupled HMC, etc.) using the bound as a stopping criterion for burn-in. Unfortunately, the second part together with the proof of the bound is in the supplementary material. The presented bound and method to compute it is, to the best of knowledge, novel and significantly extends the state-of-the-art.

After discussion, all agree that this paper makes a significant contribution and merits acceptance. These results on estimating MCMC convergence with L-lag couplings will be of broad interest to the NeurIPS community. Please take the reviewers' constructive feedback into account and follow through on your promises to improve the paper as stated in the rebuttal.

I must first admit that judging this paper was a fairly challenging task given the mixed opinions expressed by the reviewers, together with my own impressions after having scrutinized the manuscript in detail. The reviewers largely agree that the paper deserves credit as it tackles the challenging, relevant and (relatively) scarcely studied topic of restless bandit learning. I believe the main value of the paper is in the introduction of the birth-death Markov chain structure for arms of a restless bandit, together with the monotonicity and positive correlation assumptions on rewards and transitions. These are not unnatural assumptions, as evidenced by modeling literature on scheduling over wireless channels and queueing systems, and seem to greatly alleviate the computational complexity of a portion of the learning process. On the other hand, the reviewers are not fully convinced about the significance of the proposed algorithm and regret bound proven in the paper, given that the analysis is carried out for a highly structured ensemble of Markov decision processes.

Weaknesses: The paper lacks many intricate details that prevents the reader to judge the novelty and full contribution of the work. After reading the rebuttal, an overview of the proposed solution and the problem setting would be of much help to the readers. Is the entire game (with all levels) considered as a POMDP? I see sentences such as "Line 62: environment is considered as a markov process". How is the generalization problem being modelled?

Weaknesses: The work is not well presented. Terms like open-loop actions, closed-loop policies, and reachable belief space were used without definitions provided. As a result, the reviewer had difficulties understanding Figures 1 and 2. Value of information is the key of this work, but was only briefly discussed in Section 4.1. The major concern is on the evaluation of the developed methods. The POMDP community has provided a number of benchmark problems.

The authors did a good jump of addressing reviewer concerns in the response. There were some lingering concerns about whether the authors had picked the best compare-to choices for their experiments. Additional experiments and/or more careful justification for the choices made would always help. I would recommend that the authors take the reviewers' comments into account in preparing the final version of the paper.

Two out of three reviewers appreciated the contributions of this paper, with one expert reviewer praising almost every aspect of the paper. On the negative side, one reviewer took issue with the proposed setting, highlighting that the utility of the proposed objective function is somewhat dubious in the general context of multi-objective decision making. I agree with this reviewer in that having "multi-objective" in the title of the paper may set the wrong expectations for some readers, and I suggest that the authors consider changing the title of the paper for its final version to avoid such misunderstandings. Furthermore, the final version should discuss the relationship between this paper and the very recent work of Rosenberg and Mansour (2019) that studies essentially the same problem in episodic MDPs. Other than these concerns, the paper is worthy of being published without major changes.

We study regret minimization in non-episodic factored Markov decision processes (FMDPs), where all existing algorithms make the strong assumption that the factored structure of the FMDP is known to the learner in advance. In this paper, we provide the first algorithm that learns the structure of the FMDP while minimizing the regret. Our algorithm is based on the optimism in face of uncertainty principle, combined with a simple statistical method for structure learning, and can be implemented efficiently given oracle-access to an FMDP planner. Moreover, we give a variant of our algorithm that remains efficient even when the oracle is limited to non-factored actions, which is the case with almost all existing approximate planners. Finally, we leverage our techniques to prove a novel lower bound for the known structure case, closing the gap to the regret bound of Chen et al. [2021].