Thanks to all 6 reviewers for your helpful comments, we try to address all points raised in what follows: To Reviewer 1, * thank you for the positive feedback, as for your last comment regarding theoretical guarantees, this is something we are currently looking into considering, among other things, "convergent EP" of Heskes and Zoeter, 2002. To Reviewer 2, * re: "better explanation of how to find the proposal" the proposal at each node at a given step is the current approximation of the belief on that node. I.e.: the product of the approximated messages coming into that node, this is explained in line 196-209 (page 4), in particular equation 14 gives the form of the proposals. This is one of the main point of our submission i.e., suggesting to construct a proposal on the go using EP. We could try to make sure this point is clearer.

Summary and Contributions: Statistical property estimation is an important and active area at the intersection of theoretical computer science, statistics, and information theory. For example, a basic question in this realm: given n iid samples from an unknown discrete distribution p, how well can we estimate the entropy H(p), and what is an efficient algorithm for doing so? Recent efforts have shown that, for any symmetric property, the profile maximum likelihood estimator is universally minimax optimal for a wide range of parameters. While this at first seemed like a purely theoretical result, algorithmic efforts quickly caught up to show that 1) efficient approximation of the profile maximum likelihood estimator is possible and 2) approximate profile maximum likelihood estimation suffices for minimax optimality. In this context, this paper refines recent approximation algorithms from exp(-\sqrt{n} log n) to exp(-k log n) where k is the number of observed frequencies, with k O(\sqrt{n}).

This paper proposes new and substantial improvements to the algorithmic side of the PLM estimation problem. New theoretical tools are introduced and the analysis is refined and deep. The authors seem to have adequately addressed all of the concerns in the rebuttal.

A nice advantage of predictive representations of stochastic processes is that they can be expressed in terms of families of linear operators --- the "observable operators" of Jaeger (oddly, not cited in this paper; also, see Upper, and the appendix to Shalizi and Crutchfield). This paper proposes (following some earlier work) to exploit this fact, by using the instrumental variables technique from econometrics to simplify the estimation of such models. Doing so results in an estimation procedure very similar to that of Langford et al. from 2009 (reference [16] in the paper), but with some advantages in terms of avoiding iterative re-estimation. However, there seems to be an important issue which isn't (that I saw) addressed here. The instrumental variable needs to be correlated with the input variable to the regression, but independent of the noise in the regression.

This paper proposes a model for time series based on a hierarchy of sigmoid belief networks connected through time. The recently proposed Neural Variational Inference and Learning (NVIL) framework is applied to design scalable (approximate) inference and learning in the model. The model is shown to generate bouncing balls, polyphonic music, motion capture and text. Strong points - The starting point of this paper (deep directed models, variational inference) is a popular and interesting line of research. This is one paper amid several time series extensions of this family of models/inference methods and is a natural extension of this line of work.

Additional Feedback: Response to author feedback: From the informal discussion about the cross-component counters, I'm getting that it's somehow bad if different components have been explored unevenly and therefore encouraging more balanced exploration (pairwise) reduces overall variance in the amount of exploration between components. I'm sure there's a lot I'm not getting, but that helps a bit. I think it should be the case that you recover an object when you multiply its factors together (for the appropriate definition of "multiply"). There are papers (well, just one I can think of) that deal with truly factored MDPs that are the product of simpler MDPs. They correctly call their MDPs factored.

While this paper initially had some mild divergence of opinion among the reviewers, after the author response and some detailed discussion, it was agreed that this paper makes a solid contribution (please see the revised reviews). It is certainly is of relevance to NeuRIPS. After discussion, there was agreement on the significance of the conceptual contribution, namely the treatment of the cross-component bonuses. Several reviewers note that the mathematics is fairly "standard" (Bernstein-bound machinery), though in the end that should not be considered a drawback. At least one reviewer notes that the 31pp appendix means that it is not possible to verify the mathematical results during the review period.

Summary and Contributions: The paper proposes a method for utilizing ODEs to represent dynamics for continuous-time decision-making problems with the aim of They also target filling a perceived gap in the literature of Deep RL for continuous-time problems, where most publications are model-free and discretize time if it is continuous. They claim that their approach leads to lower dependence on vast amounts of training data, better performance and that the model-based approach is well-founded. I tend to agree, although this is not exactly my area. I also believe the importance of connecting ODEs and other explicit models is critical for extending RL methods to important problems in physics, chemistry, epidemiology and population modelling.

The reviewers all found the paper to make a reasonable contribution. While there are some limitations that were pointed out, the rebuttal addressed them well and the experiments were appreciated.

SUMMARY Hamiltonian MCMC methods sample from a probability distribution by treating its log as a "potential energy" function over the state space, augmenting the space with extra "momentum variables" and their associated "kinetic energy", and evolving the state of the Markov process by integrating the physical Hamiltonian equations of motion of the system. Each step of the Markov chain is accomplished by numerically integrating the Hamiltonian equations forward in time. However, if the energy function is non-differentiable, the integral is not well-defined. The rejection step that is used to counteract numerical inaccuracies in the integration also accounts for such non-differentiable regions, but at the cost of slowing down the mixing rate of the Markov chain. This paper suggests physically-inspired "reflections" and "refractions" of the trajectory of the system that occur whenever the state crosses a discontinuity in the energy function. It applies to target distributions that are differentiable everywhere except on the boundaries of certain polytopes; the reflection or refraction occurs whenever the trajectory of the system crosses such a boundary.