NeurIPS 2020

An Equivalence between Loss Functions and Non-Uniform Sampling in Experience Replay


Review 1

Summary and Contributions: In this paper they argue that there are two different contributions to PER. On the one hand there is the change in the direction of the gradient based on the prioritization schemes and on the other hand there is the non-uniformity of the updates themselves (which relates to the rate of signal propagation throughout the TD learning). They show a connection to changing the loss itself and leaving the sampling scheme uniform and perform multiple experiments comparing the various methods.

Strengths: The paper is well written and is a novel approach. Specifically, the ability to map a priorized sampling scheme to an equivalent (in expectation) gradient with respect to a uniform sampling scheme may drive many future research directions and innovations.

Weaknesses: While in expectation the gradient of PAL and LAP are identical, the behavior in complex domains is not. Investigating this question is very important and is left as an open hole in the work. One explanation that could hold is that although the gradient is identical in expectation, the prioritization scheme causes the signal to propagate faster. In this case, the difference between LAP and PAL may be tied to the periodicy of the MDP. MuJoCo domains are very simple and repetitive, thus random uniform sampling is probably sufficient for signal propagation and in the ATARI domains the prioritization provides a large benefit.

Correctness: Yes.

Clarity: Very.

Relation to Prior Work: Yes.

Reproducibility: Yes

Additional Feedback: I enjoyed reading this paper. The idea is very interesting and novel. Very simple to grasp, very simple to implement. I believe this work can have a real impact (in terms of adoption) on the community. ** post rebuttal ** I really liked this paper. I think that it can be improved by looking into the effects of each method based on the periodicity of the MDP.


Review 2

Summary and Contributions: This paper analyzes the loss function induced by prioritized experience replay (PER), and finds that it is possible to remove the importance sampling ratio in the loss used with PER by instead using a different loss altogether. The theoretical analysis provides a more general analysis of what losses can be used when sampling from a distribution different than uniformly sampling the replay buffer. This analysis leads to the paper presenting a new prioritization scheme, which they call LAP, and a new loss, which is referred to as PAL. Experiments compare these two new schemes to methods that use a replay buffer in both the continuous control and atari benchmarks.

Strengths: * A deeper analysis of methods that have been found to be empirically useful is of immense value to the community. In this case, the perspective of considering the loss function that PER induces and then attempting to remove the requirement for an importance sampling ratio are commendable. * Further breaking down the benefit of PER as variance reduction due to non-uniform sampling and changing the expected gradient due to the induced loss function is a useful insight. * Encouraging Experimental results

Weaknesses: Section 5 was harder to follow than the rest of the paper. If I understand correctly, the LAP procedure aims to use the Huber loss after sampling non-uniformly rather than using the L1 loss. In order to make this loss unbiased the LAP procedure suggests a new prioritized sampling procedure. PAL then proposes a loss function to be used if we want to to sample from the buffer uniformly but would like to have the same expected gradient as LAP. This idea could be put forth a little more clearly.

Correctness: There do not seem to be any glaring issues in the correctness of the theoretical results. **After Rebuttal** I am satisfied with the empirical methodology for the most part. In the experiments on continuous control, the separation in the performance of SAC and the proposed techniques happens just barely at the point where they are evaluated. Perhaps a better method to compare the two techniques is to look at the area under the learning curve or evaluation at various points in learning. Authors did not respond to this suggested metric.

Clarity: I have some issue with the clarity of the paper. Various terms are used without previous introduction. Examples are the horizon T and target network parameters \theta' in Section 3. Another ambiguous statement is the summation when computing p(i) just before Equation (6). The summation in the denominator is not clear about what range it is summed over. I assume it is batch D, but would be helpful to be clear. Another clarification is whether \delta(j) in the denominator should also be an absolute value. **After Rebuttal** Authors have agreed to look into above points.

Relation to Prior Work: Relation to previous work is discussed sufficiently

Reproducibility: Yes

Additional Feedback: The paper suffers in its readability and clarity. See comments above. Additional discussion with the other reviewers and AC led to the following comments: 1. Corollary 1 seems not quite correct. They're applying the fundamental theorem of calculus so that a gradient and an integral cancel out. But this requires that the function inside the integral is continuous. While this is quite mild, the function inside is itself a gradient! So if we had a ReLU network, the gradient of the loss would be highly discontinuous. 2. Corollary 1 is also stated in a weird way. In Theorem 1, the conclusion is that L_1 and \lambda L_2 have the same expected gradients. But because they have 1/\lambda in the expression in Corollary 1, you actually find that L_1 and L_2 themselves have the same expected gradients. So formally Theorem 1 is not satisfied by this construction. They should either remove the 1/\lambda or change Theorem 1.


Review 3

Summary and Contributions: This papers examines prioritized experience replay, where transitions are sampled in a non-uniform manner to constitute the optimization objective, and presents a new uniformly sampled loss function that yields the same gradients in expectation. The paper provides a theoretical derivation and also shows promising empirical results that back up the claims. Update: Thanks to the authors for their clarifications!

Strengths: * Novel theoretical contribution of demonstrating why Prioritized Experience Replay works * A new loss function (loss-adjusted PER) and sampling algorithm (PAL and LAP) for updating RL agents * Promising empirical results

Weaknesses: * No major weaknesses. If I were to point out something, it would be that the paper could perhaps use some empirical analyses (beyond learning curves on different environments), probing what conditions or assumptions benefit PAL and LAP over traditional PER.

Correctness: * Claims seem largely correct to me

Clarity: * The paper is very well written! I appreciate the way the connection between sampling and loss functions was established in Section 4 before deriving the new schemes in section 5.

Relation to Prior Work: * yes, related work is well represented.

Reproducibility: No

Additional Feedback: * It would help to provide details on exact models, training schemes, etc. used for reproducibility. Releasing the code would be even better. * Line 6 (abstract): The sentence leading with 'surprisingly' put me off a bit in the abstract without further information that comes later in the paper -- if the theory is showing they are equivalent, why should the empirical results working out be surprising?


Review 4

Summary and Contributions: This paper argues that original PER has bias and it requires importance sampling. In order to solve this problem, the authors derive an unbiased loss condition for a given priority scheme. From this derivation, the authors suggest an unbiased prioritization method without importance sampling by using the Huber loss (LAP), and loss with uniform sampling, which is equivalent to LAP (PAL).

Strengths: The paper suggests a prioritization method with unbiased loss for given priority scheme and converts it to equivalent loss with uniform sampling from the buffer. Furthermore, LAP considers a minimum variance criterion when TD error is large (L1 loss). By applying LAP to TD3 and DDQN, the performance is enhanced notably. Also, PAL uses equivalent loss with uniform sampling, which has the same expectation of loss gradient. Thus, it does not require complex prioritization scheme maintaining the advantage of prioritization.

Weaknesses: The paper argues that PAL can avoid prioritization scheme by using equivalent loss, but the performance of PAL is much lower than that of LAP. As explained in the paper, LAP considers the minimum variance criterion, but PAL cannot reflect the criterion. Thus, PAL cannot completely replace LAP and one still has to consider prioritized sampling from the buffer to have advantage of prioritization. LAP enhances the performance of baseline algorithms based on unbiased estimation and variance reduction, but the enhancement is limited for specific tasks. For example, the performance of other tasks except Humanoid in continuous domain does not increase much. That is probably because TD3 already works well in most tasks, so it seems to be necessary to do additional experiments.

Correctness: The suggested prioritization method is correctly unbiased without importance sampling and the empirical methodology is also correct.

Clarity: The overall flow of the paper is well written. The proof of theorem is easy to understand and several observations and corollary help understanding.

Relation to Prior Work: It gives clear description for the suggested prioritization method and the difference from the existing PER method.

Reproducibility: Yes

Additional Feedback: ====== After author response ====== The authors wrote the feedback well, and I understood and accepted it.