NeurIPS 2020

Reinforcement Learning with Augmented Data


Review 1

Summary and Contributions: This paper presents Reinforcement learning with Augmented Data (RAD), a plug-and-play module in reinforcement learning to naturally apply data augmentation. The experiments show that using RAD can improve the data efficiency and final performance on both pixel-based environments in Deepmind control suite and state-Applying domain knowledge is quite important to reinforcement learning. RAD can help learning the representation in reinforcement learning, which might be difficult before. This paper also systematically studies data augmentation to pixel-based environments, analyzing which kinds of augmentation are the most important. The learned representation seems to work quite well and can also generalize. The code is released and detailed instructions are attached, which greatly helps reproducibility. based Mujoco environments. RAD can also improve the test-time generalization on pixel-based environments (OpenAI ProcGen). The paper also studies which data augmentation works better and proposes two new data augmentation.

Strengths: Applying domain knowledge is quite important to reinforcement learning. RAD can help learning the representation in reinforcement learning, which might be difficult before. This paper also systematically studies data augmentation to pixel-based environments, analyzing which kinds of augmentation are the most important. The learned representation seems to work quite well and can also generalize. The code is released and detailed instructions are attached, which greatly helps reproducibility.

Weaknesses: The experiments on state-based environments are not convincing enough to me. I have some questions on the reported numbers. It seems that the reported performance of TD3 is quite bad, compared to TD3 paper. For example, TD3 can achieve ~7k rewards in HalfCheetah-v1 after 200k steps according to the TD3 paper, while the reported number here is only 3k. In Walker2d, TD3 paper reports ~2k total rewards after 200k steps, while the number here is < 0. Why is the difference so large? It seems that the error bars are also very large. How many seeds are used to evaluate each algorithm?

Correctness: It's hard to read Figure 12. Maybe smoothing the curves can help. Figure 12 only plots the curves for at most 200k steps, which are not enough to conclude the superior final performance of using RAD. Also the figure prevents me from finding a good gap between SAC and SAC+RAD. L261-263: If multiple variables are used, the relative differences can also be changed. So this argument can't explain why random amplitude scaling can help. Also random amplitude scaling can hurt in some scenarios, e.g., the state is a scalar.

Clarity: In general, this paper is written well. Some minor comments: 1. L259: nothing -> noting 2. As I mentioned above, Figure 12 is difficult to read.

Relation to Prior Work: Authors discuss the difference between RAD and prior work in Section 2 and Appendix A.4.

Reproducibility: Yes

Additional Feedback: POST REBUTTAL: I've read other reviews and authors' rebuttal. The authors answered my questions but I'm still not very convinced why TD3 performs so bad. I'd like to keep my score.


Review 2

Summary and Contributions: This paper investigates data augmentation in RL. Its main contribution is showing that data augmentation is in fact very powerful for RL and could achieve SOTA on some envs.

Strengths: The paper is generally well written and well prepared with extensive related work and code. The empirical study is very important and the results are to some extent significant. The paper should be interesting for all RL researchers.

Weaknesses: While indeed to my knowledge there hasn't been an extensive study of data augmentation in RL (data augmentation, and some other deep learning techniques are curiously missing in RL), the results in this paper are somewhat expected / not that shocking. Data augmentation is very effective on pixel-based RL since they are originally from computer vision. So the SOTA performance of RAD on pixel-RL is more because it helps convolutional neural nets than anything to do with RL. And results on state-based RL (e.g. OpenAI Gym) are not that significant. In the end, it's still that data augmentation helps CNN on images, no matter it's RL or supervised learning. Still, I think it's a good paper and should be published. It has its significance in pixel-based RL itself with many recent works (CURL, DREAMER, PlaNet, etc). In the meantime, due to my comments above, the results are indeed only to some extent significant, so I don't think it warrants a higher score. Even though some augmentation techniques are claimed to be new, the method part itself presents little, if not none, novelty. The novelty could be argued to be demonstrating the potential of data augmentation in RL. One problem on writing: "random translation" is claimed to be one of the two newly introduced augmentations in abstract and introduction, but it's never specifically introduced in Methods (Sec.4), and in Sec.4 the two new techniques proposed are "random amplitude scaling" and "Gaussian noise". This is a paper that readers can grasp in less than five minutes, so there shouldn't be mistakes like this, which must be corrected before publication.

Correctness: Generally yes.

Clarity: Mostly yes.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: ===== post rebuttal ===== Just be sure to correct the confusion I've mentioned.


Review 3

Summary and Contributions: The paper investigates empirically the application of various forms of data augmentation to state-of-the-art reinforcement learning algorithms. Two novel forms of data augmentation are proposed for state-based RL. The empirical verification on a wide variety of RL benchmark problems with continuous control inputs overwhelmingly demonstrates the benefit of data augmentation with respect to reducing significantly the data complexity of the original methods.

Strengths: A major advantage of the proposed data augmentations is that they do not require any modifications to the baseline RL algorithms, and their benefits can be realized by simply appending the original observations with the augmented data.

Weaknesses: Although the proposed general method looks very useful, there does not seem to be a major algorithmic advance in this paper. Most forms of data augmentation used in the evaluation have been known from before, with the exception of the two augmentations specific to state-based RL. Moreover, I am not entirely sure why the data augmentation proposed in this paper works, and this paper does not provide a very good explanation. The main idea is to use data augmentation techniques already used in the area of image classification. However, whereas in image classification problems the class label is invariant to translations, rotations, partial occlusions, some color variations, etc., in RL the correct action in a particular state, respectively when seeing the observation in that state, should not be invariant to them. If there is a one-to-one mapping from state to observation, wouldn't data augmentation make this mapping one-to-many, thus making policy learning harder? I understand that the augmentation is added to the original observation, instead of replacing it, but still I do not understand why it helps. The argument that randomly scaling the true state variable would make the algorithm robust to noise is perhaps acceptable, but all of these simulation environments are deterministic, so this argument should not apply in this verification.

Correctness: They appear to be correct.

Clarity: The paper is written very well.

Relation to Prior Work: The paper does a very good job in reviewing prior work and discussing similarities to and differences from it.

Reproducibility: Yes

Additional Feedback: Minor typos: P.4, L.132: "may perceived" -> "may be perceived" P.6, L.200: "in in" -> "in" --------------------------------------------------------------------------------------------------- Post rebuttal: I think the authors addressed my concerns satisfactorily, although, in the end, there is no theoretical insight as to why data augmentation works, and the main argument for using it are the empirical results presented in the paper. In spite of this, I still think that this paper should be accepted - it proposes a trick that appears to be useful, and does not take a whole lot of effort to implement, so it could be a very useful and practical addition to the state of the art that RL practitioners might appreciate.


Review 4

Summary and Contributions: This paper proposes a reinforcement learning method by incorporating data augmentation into RL. The main contributions that introduce two data augmentations: random translation and random amplitude scaling. Although the author said that it is the first extensive work for RL with data augmentation, the proposed method lack of novelty, because it looks like borrow some augmentation tricks from computer vision and apply it on RL.

Strengths: + The experiments are sufficient and the figures are verifying the data augmentation is working on RL. + The work is well written and easy to read. + Enable a new method for RL with data augmentation. + A structured system. + Reasonable experiments outperforming prior work in qualitative and quantitative evaluations.

Weaknesses: - The method is simple that just like combining data augmentation with RL. - The novelty is not clarified. In summary, the contributions are not enough, which includes some data augmentations. - The work looks nice and well written. However, the technical components adopted in the work are most from the existing ones. It's more a engineering work for a nice application. - We expect to see the challenges as generating images.

Correctness: Yes.

Clarity: Good.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: The feedback addresses my concerns. I hope the authors should provide more details about the novelty in the revision.