Prediction and Control in Continual Reinforcement Learning

Temporal difference (TD) learning is often used to update the estimate of the value function which is used by RL agents to extract useful policies. In this paper, we focus on value function estimation in continual reinforcement learning. We propose to decompose the value function into two components which update at different timescales: a permanent value function, which holds general knowledge that persists over time, and a transient value function, which allows quick adaptation to new situations. We establish theoretical results showing that our approach is well suited for continual learning and draw connections to the complementary learning systems (CLS) theory from neuroscience. Empirically, this approach improves performance significantly on both prediction and control problems.

Paper

References (52)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer mEyq7/10 · confidence 3/52023-06-28

Summary

In this work, the author proposes a new continual reinforcement learning method, the PT-TD, which consists of two value networks, one learns the transient memories and the other learns the permanent memories. The author proves that in the semi-continual setting, the PT-TD algorithm will get better performance and lower forgetting, and the same result can also be found in the continual reinforcement learning setting.

Strengths

The algorithm of this article is well-proven and well-demonstrated. The author not only proves the PT-TD algorithm in a mathematical way but also gives out detailed and rigorous experiments. Also, the article is really well-written and easy to understand. All of the hyperparameters are studied. Meanwhile, it is appreciated that the author discussed semi-continual reinforcement learning and continual reinforcement learning separately.

Weaknesses

First, even though the proof is very well, the algorithm itself is not very groundbreaking, or, the result is not very rewarding. Besides, the environments of the experiments are simple.

Questions

I think the formulation of the Theorem 8 is a little confusing. Maybe you can give out explicitly which network is using (P or T). Can your algorithm generalize to continuous action space? How should I learn the actor network at this time? I notice that in the continual reinforcement learning setting, the learning rate of T network is always higher than it of P network. According to [1], if it is the difference in the learning rate affect good performance? [1]Arani, E. et al. “Learning Fast, Learning Slow: A General Continual Learning Method based on Complementary Learning System.” ArXiv 2021

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

4 excellent

Presentation

3 good

Contribution

2 fair

Limitations

None

Reviewer zZwF7/10 · confidence 3/52023-07-05

Summary

This paper proposes new algorithms for prediction and control in continual reinforcement learning. The new algorithms use permanent and temporary memories. The permanent memory captures a general value function for all the tasks, while the transient memory is used to adapt to the current tasks continually. The paper provides theoretical results for the prediction setting with a tabular learner. Additionally, the paper provides empirical results showing that their method outperforms vanilla TD-based algorithms in various environments using different types of function approximators.

Strengths

The paper directly tackles the continual reinforcement learning setting. This is a challenging problem setting with few general algorithms and evaluation environments. The paper provides prediction and control algorithms for the continual RL setting. Theoretical proofs of convergence are provided for the prediction setting with a tabular learner. The extensive and sound empirical investigation shows that the proposed method can generally outperform existing methods in various environments and with different various function approximations, including Deep Networks.

Weaknesses

The paper does not have any major weaknesses. I did not check the proofs, so I can not comment on the correctness of the proofs. However, there are a few minor issues that can be fixed to improve the paper. 1. I'm confused about Equation 5, specifically line 112. Does it mean the "gradient" or the "semi-gradient" of the transient memory? 2. What are the confidence intervals being reported? Line 222 says it is the 90% confidence interval, but is the bootstrapped confidence interval or confidence interval based on the standard error? And if it is based on the standard error, how do you know it is the 90% confidence interval? 3. The continual deep RL algorithms used in the paper use standard deep RL tools like Experience replay and target networks. However, the exact use of these tools is not fully explained in the paper (or the appendix). This creates a lot of questions; for example, are there a total of three networks, target, transient and permanent? I suggest that the authors include a section in the appendix that gives the pseudocode of the PT-DQN algorithm. The lack of details about the PT-DQN algorithm makes it hard to replicate the results given in the paper.

Questions

1. Line 88 says, "... multi-task RL methods ..." What is multi-task RL, and how is it different from continual RL? What limits the use of multi-task RL methods in continual RL? The second question should be answered in the main text. 2. Figure 4d shows the performance of different versions of DQN on Minigrid. Why is the performance of all the methods, particularly DQN(reset), so much worse during the second task? If the DQN is being reset at each task, it should perform similarly on Episodes 500 and 1000. Is this because the first goal state in all the runs is the same? 3. What is PT-DQN-0.5x in Figure 5? 0.5x should be described in the main paper. 4. Lines 733-742 describe the training details for the deep RL experiments. However, many of these choices seem arbitrary, or the paper does not explain the reasoning behind these choices. Why are the conv layers not updated when updating the permanent memory? Why isn't there a fully independent permanent value network? Wouldn't that be closer to the algorithms presented in the paper? 5. Lines 803-805 say that the permanent memory was trained with SGD while the rest of the network was trained with Adam. Why this specific choice? I'm guessing the authors also tried Adam to update the permanent memory. Why didn't Adam work for permanent memory? 6. The plots (like Figure C.13) showing the sensitivity curve will be better with a log scale for the x-axis. 7. Why isn't there any effect of PM LR on the performance in the minigrid environment (Figure C.13c)?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

One limitation of this work is that it only proposes methods for value-based RL. It is unclear how the idea of permanent and transient memory can be extended to policy-gradient based RL. A discussion on this direction of future work will be helpful.

Reviewer 27Kd6/10 · confidence 4/52023-07-05

Summary

The paper proposes a method to improve continual learning in reinforcement learning with non-stationary dynamics. It proposes to maintain two value functions: one that tracks information about the current task (transient) and one that maintains a general knowledge of the task distribution (permanent) in order to better tackle the non-stationarity of the environment.

Strengths

1. The general idea is simple in terms of intuition and implementation. I like the idea of maintaining two different value functions operating different time-scales. 2. Assuming soundness (see below), the experiments well-done and are insightful.

Weaknesses

While the overall approach is intuitive, the specifics dont seem very intuitive to me and there is not much information given about it. For example, Eq 4 (see questions below) is not intuitive to me on how it was derived. Similarly, I have other questions.

Questions

1. What do the authors mean when they say prior work “reflected the nature of that task” in line 41? 2. What is the intuition for the update in Eq 4? Why does the update to the parameters of the P get scaled by V^PT? Similarly, what is the intuition for Eq 5? Why is \delta purely in terms of the overall update, PT? 3. The definition in Eqn 2 and 3 is that both memories are equally weighted and are just a sum. I am curious what the authors think about this equal weighting and the consequences of it? 4. It is a little unclear to me why only the initialization condition is sufficient for Theorem 1 to hold? 5. Shouldn’t the equation in Theorem 2 be what is in Equation 5 i.e. PT instead of P? 6. It may be worthwhile to see how Continual backprop [1] can be incorporated into the proposed method. [1] Continual Backprop: Stochastic Gradient Descent with Persistent Randomness. Dohare et al.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See above questions

Reviewer mFKc4/10 · confidence 3/52023-07-06

Summary

This work proposes a reinforcement learning method for prediction and control. The value function in reinforcement learning is decomposed into two terms: a permanent value function, and a transient value function. The permanent function contains the long term knowledge of the interplay between the agent and the environment, and the transient value function absorbs the quick information variations. This work uses temporal difference learning for updating the estimate of the value function estimation. The learning process is related to the complementary learning systems from neuroscience. Prediction experiments are performed for 5x5 discrete grid. The agent moves from the center and selects an action the goal states.

Strengths

It is of great interest to extract valuable information from stream data. The development of organic transistor featuring a selective transition from short-term to long-term plasticity of the biological synapse. Catastrophe forgetting is a major challenge for machine learning. A possible solution for solving such a problem is very attractive. It is closely related to meta learning.

Weaknesses

The idea for separating long term and transient information is intuitive. It has been used in many machine learning algorithms such as LSTM. It is not clear if a long term function is always available, especially for complicated interplay between an agent and its environment. The separation of long term and transient can be dynamic. The settings for the experiments are very simple. It does not demonstrate the possibilities for real applications.

Questions

In the experiments, the proposed method is compared with DQN. Is it possible to compare other competitive algorithms?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

The limitations of this work are not explicitly given.

Reviewer mEyq2023-08-14

Thanks to the author for the rebuttal. Just to further refine this paper, I propose to add experiment results of the setting that both P and T networks use the same learning rate.

Authorsrebuttal2023-08-16

Thank you for your suggestion. We tested our method on a broad range of learning rate values including the cases where they are the same for both the P and T networks. We picked the combination that resulted in the smallest error (or the largest reward) for each experiment. We included a detailed report on the hyperparameter selection in the appendix (Section C4). In summary: (a) Our approach is less sensitive to learning rate choices when compared with its vanilla counterparts; (b) The performance degraded slightly when we used large learning rate values to train the permanent memory (this is expected because we want the permanent memory to take small steps to allow slow knowledge consolidation). We will elucidate our observations in the main paper based on our discussion, and include a reference to the suggested paper.

Reviewer zZwF2023-08-16

I thank the authors for answering my questions. I maintain my score and recommendation for acceptance.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC