Update: ----------- I had a look at the author response: It seems reasonable, contains a lot of additional information / additional experiments which do address my main concerns with the paper. Had these comparisons been part of the paper in the first place I would have voted for accepting the paper. I am now a bit on the fence about this as the paper could be accepted but will require a major revision, I will engage in discussion with the other reviewers and ultimately the AC has to decide whether such big changes to the experimental section are acceptable within the review process. Original review: --------------------- The paper presents a method for transfer learning via a variational inference formulation in a reinforcement learning (RL) setting. The proposed approach is sound, novel and interesting and could be widely applicable (it make no overly restrictive assumptions on the form of the learned (Q-)value function).

This paper is already available on arxiv and cited 10 times. It is a very good paper introducing a new approach to count-based exploration in deep reinforcement learning based on using binary hashcodes. The approach is interesting, the presentation is didactical, the results are good and the related literature is well covered. I learned a lot from reading this paper, so my only criticisms are on secondary aspects. I have few global points: - I think it would be interesting to first train the convolutional autoencoder, and only then use your exploration method to compare its performance to state-of-the-art methods.

The authors propose to augment value-based methods for deep reinforcement learning (DRL) with batch methods for linear approximation function (SRL). The idea is motivated by interpreting the output of the second-to-last layer of a neural network as linear features. In order to make this combination work, the authors argue that regularization is needed. Experimental results are provided for 5 Atari games on combinations of DQN/Double DQN and LSTD-Q/FQI. Strengths: I find the proposition of combining DRL and SRL with Bayesian regularization original and promising.

The paper considers the problem of online learning in two-player zero-sum stochastic games. The main result is constructing a strategy for player 1 that guarantees that the cumulative rewards will never go below the maximin value of the game by more than a certain bound, no matter what strategy the other player follows. The bound is shown to grow sublinearly in the number of rounds T of the game, and polynomially on other problem parameters such as the diameter, the size of the state and action spaces. The results imply that the proposed algorithm can be used in self-play to compute near-maximin strategies for both players. The algorithm and the analysis are largely based on the UCRL algorithm of Auer and Ortner (2007) and the analysis thereof.

Posterior Sampling for Reinforcement Learning: Worst-Case Regret Bounds This paper presents a new algorithm for efficient exploration in Markov decision processes. This algorithm is an optimistic variant of posterior sampling, similar in flavour to BOSS. The authors prove new performance bounds for this approach in a minimax setting that are state of the art in this setting. There are a lot of things to like about this paper: - The paper is well written and clear overall. I would say that most of the key insights do come from the earlier "Gaussian-Dirichlet dominance" of Osband et al, but there are some significant extensions and results that may be of wider interest to the community.

The manuscript discusses an important topic, which is optimization in deep reinforcement learning. The authors extend the use of Kronecker-Factored approximation to develop a second order optimization method for deep reinforcement learning. The optimization method use kronecker-factored approximation to the Fisher matrix to estimate the curvature of the cost, resulting in a scalable approximation to natural gradients. The authors demonstrate the power of the method (termed ACKTR) in terms of the performance of agents in Atari and Mujoco RL environments, and compare the proposed algorithm to two previous methods (A2C and TRPO). Overall the manuscript is well-written and to my knowledge the methodology is a novel application to Kronecker-factored approximation.

It is well known that the standard TD algorithm widely used in reinforcement learning does not correspond to the gradient of any objective function, and consequently is unstable when combined with any type of function approximation. Despite the success of methods like deep RL, which combines vanilla TD with deep learning, theoretically TD with nonlinear function approximation is demonstrably unstable. Much work on fixing this fundamental flaw in RL has been in vain, till the work on gradient TD methods by Sutton et al. Unfortunately, these methods work, but their analysis was flawed, based on a heuristic derivation of the method. A recent breakthrough by Liu et al. (UAI 2015) showed that gradient TD methods are essentially saddle point methods that are pure gradient methods that optimize not the original gradient TD loss function (which they do not), but rather the saddle point loss function that arises when converting the original loss function into the dual space.

This paper presents a RL optimization scheme and a theoretical analysis of its transfer performance. While the components of this work aren't novel, it combines them in an interesting, well-presented way that sheds new light. The definition of transfer given in Lines 89–91 is nonstandard. It seems to be missing the assumption that t is not in T. The role of T' is a bit strange, making this a requirement for "additional transfer" rather than just transfer. It should be better clarified that this is a stronger requirement than transfer, and explained what it's good for -- the paper shows this stronger property holds, but never uses it.

Summary: "A Unified Game-Theoretic Approach to Multiagent Reinforcement Learning" presents a novel scalable algorithm that is shown to converge to better behaviours in partially-observable Multi-Agent Reinforcement Learning scenarios compared to previous methods. The paper begins with describing the problem, mainly that training reinforcement learning agents independently (i.e. each agent ignores the behaviours of the other agents and treats them as part of the environment) results in policies which can significantly overfit to only the agent behaviours observed during training time, failing to generalize when later set against new opponent behaviours. The paper then describes its solution, a generalization of the Double Oracle algorithm. The algorithm works using the following process: first, given a set of initial policies for each player, an empirical payoff tensor is created and from that a meta-strategy is learnt for each player which is the mixture over that initial policy set which achieves the highest value. Then each player i in the game is iterated, and a new policy is trained against policies sampled from the meta-strategies of the other agents not equal to i.

Post discussion update: I have increased my score. In particular they took to heart my concern about running more experiments to tease apart why the system is performing well. Obviously they did not run all the experiments I asked for, but I hope they consider doing even more if accepted. I would still like to emphasize that the paper is much more interesting if you remove the focus on SOTA results. Understanding why your system works well, and when it doesn't is much more likely to have a long-lasting scientific impact on the field whereas SOTA changes frequently.