NeurIPS 2019
Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
Paper ID:7793
Title:Generalization in Reinforcement Learning with Selective Noise Injection and Information Bottleneck

Reviewer 1

This work builds on the previous work about generalization in RL ([10] in the paper references) by (re-)investigating the classical stochastic regularization approaches in this context. It completes and updates the claims made in [10] by focusing of similar performance based experiments. Clarity: The method is clearly described in the paper. Significance: The question of generalization in RL is of great interest to the field. Main comments: - The paper motivates well the problems one faces when is comes to regularization in RL. The idea of being “selective” when injecting a regularizing noise makes a lot of convincing. However, the proposed method is more about an ad-hoc mixture of training modes, with noise injection and without, which makes the notion of selectivity here a bit loose. - Related to the previous point, since the mixture can be done in several ways, a comparison with other ways in order to justify this choice of the mixture can greatly enhance the paper quality. For example, one can imagine adapting the dropout probability or scheduling it in a curriculum learning fashion. How does your approach relate/compare to this ? - In the multiroom environment, it is confusing that your method has the highest variance while the method was presented as to be favoring variance reduction. Also, IBAC seems to have a higher trend than IBAC-SNI above 0.4 x 10^8 frames (IBAC has not stabilized like IBAC-SNI), which makes the fact that IBAC outperforms IBAC-SNI very likely when trained for longer than 10^8 frames. - It is not clear from eq. 7 how SNI circumvents some of the issues listed as motivation for it. For example, how does the roll-out avoid “undesirable actions (potentially ending episodes prematurely) ? - From the experiments point of view, a deeper analysis could be expected to relate to the issues listed as being specific to regularization in RL. The variance of the terms described in section 3, the nature of the trajectories ... etc ? In other words, it is not clear how SNI deals with the stated issues in section 3. Minor comments: - Should the normalization constant Z disappear in the right hand side of eq. 8 ? Also does this equation correspond to the semi-gradient or the exact gradient ? - Was L^V_{AC} defined in the paper ? and V_\theta(z) ? (I might have missed them) - In figure 3, can we assume that the curves with no std have very small variance ? Does the averaged curves over more runs for those one change the observations ? - When you say “outperforms an equal mixing = 0.5 early on in training ”, to which curve are you relating here ? Is it Dropout with \lambda=0.5 ? ------------------------------------------------------------------------ I have read the author response which has fairly addressed most of my concerns so i decided to increase my score. I would expect the authors to introduce more convincing plots of the claimed performances (like the one in the rebuttal) and a clearer justification of their choice of this type of mixture, in a future version of the paper.

Reviewer 2

The main contribution of the paper is to introduce a novel usage of the Information Bottleneck technique for off-policy actor-critic algorithms involving importance sampling and function approximation based critics, which results in state of the art results on the CoinRun challenge domain. An insightful discussion on some of the challenges when applying techniques that improve the generalization of supervised learning techniques to some RL settings. In general I think the paper is well motivated and supported sufficiently with experimental evidence. I have a couple of presentation concerns but generally think the paper is acceptable. Presentation: In general the paper is clearly written, however I do think there is a fundamental misunderstanding in the presentation, and think the authors would do well to avoid statements like "in RL the loss function depends on the learned parameters" or phrases like "the RL loss". In IID supervised learning, while one typically talks about having the same loss function (such as square loss) for all examples, with a relatively simple form, the loss function itself is a function of the parameters, so what is said isn't strictly correct. It's overly simplistic to view every single RL algorithm or technique as minimizing some loss function. Some algorithms are expressed in the form of weight update mechanisms, and rely on being contraction mappings or what not to converge, with function approximation being applied heuristically to scale to larger problems. Policy gradient techniques estimate gradient of the expected future return under some parametrized policy, it's not a loss function per-se, it's a goal you can estimate a gradient with respect to and make local improvements. Perhaps I am being overly pedantic, but I think the presentation would be more clear if you explicitly state the setting you are interested in (the empirically successful class of off-policy actor-critic algorithms involving importance sampling and function approximation based critics) and limit your broader statements and intuition to this setting. Too be clear, I have no problem with the motivation or insights behind the algorithmic changes, I just think the broader discussion is a bit sloppy. More minor points: - In line 66, the expectation should be over q not m. - Use q(\cdot) to talk about the distribution q, not q(m) which is a density. Line 64, just write m ~ q. Similarly for other distributions in the paper. - I don't understand the comment on line 233 - isn't the mean performance ~0.4? Do you mean that one seed got to a return of 1? - I think the paper could be improved too if it is possible to run larger numbers of seeds. The stderr is somewhat meaningless for such small sample sizes like 5.

Reviewer 3

Clarity The paper is clearly written. The main ideas, experiments, and results are described clearly. Originality The identification of noise issues and the proposed solutions appear original. The idea of selective noise injection is new to me, and it addresses the issue with noisy gradients. Although an information bottleneck has been proposed in the past, its evaluation in this setting seems novel. Significance This work should be of use to the deep reinforcement learning community. Techniques such as batch normalization from supervised deep learning do not perform consistently in deep reinforcement learning. This paper's examination of several regularization methods on Coinrun in section 5.3 provides additional evidence of this phenomenon, and the paper provides potentially useful alternatives. Quality The technical aspects of the paper are acceptable. The issues are identified, and the proposed solutions are well motivated. The experiments and evaluation is clear. The presentation could be improved in a few ways as noted below. line 71: Please don't use r_t for the importance sampling ratio in Equation 1 when it also means the reward in equation 5. line 95: Was dropout only applied on the last layer in all the experiments in this paper? Equation 5: The rollout policy should be evaluated at s_t not s_T Equation 12: The number of parameters that are aliased or underspecified is confusing. I see lambda (equation 7), lambda_H, and lambda_V. I also see L^V_{AC} without a definition, but it is perhaps equation 6. However the IBAC algorithms are referenced without a lambda_H or lambda_v term in the experiments Figures 2,3). Please clarify the use of these terms.