Gradient Informed Proximal Policy Optimization

We introduce a novel policy learning method that integrates analytical gradients from differentiable environments with the Proximal Policy Optimization (PPO) algorithm. To incorporate analytical gradients into the PPO framework, we introduce the concept of an α-policy that stands as a locally superior policy. By adaptively modifying the α value, we can effectively manage the influence of analytical policy gradients during learning. To this end, we suggest metrics for assessing the variance and bias of analytical gradients, reducing dependence on these gradients when high variance or bias is detected. Our proposed approach outperforms baseline algorithms in various scenarios, such as function optimization, physics simulations, and traffic control environments. Our code can be found online: https://github.com/SonSang/gippo.

Paper

References (35)

Scroll for more · 23 remaining

Similar papers

Peer review

Reviewer 39SH5/10 · confidence 3/52023-06-26

Summary

This paper studied the combined use of both the analytical policy gradient and the likelihood ratio policy gradient for training policy networks, based on the PPO algorithm. To make the combination feasible, a new alpha-policy is introduced and its approximation technique has been successfully developed. Besides some theoretical studies, empirical results further show the potential usefulness of the newly proposed algorithm in solving some benchmark problems.

Strengths

The theoretical study in this paper regarding the analytical policy gradient and the alpha-policy sounds interesting and novel. The experiment results show that the new algorithm can be very useful on some benchmark problems.

Weaknesses

While it is interesting to mix analytical policy gradients with learned policy gradients to enhance the reliability and performance of the policy network training algorithms, this idea is clearly not restricted to on-policy algorithms such as PPO. Although the authors made it clear that PPO is one of the most popularly used on-policy algorithms, it remains questionable why this paper only studies the effectiveness of using combined policy gradients in PPO. The possibility and potential limitations of using the proposed policy gradient combination technique on other algorithms, particularly off-policy algorithms, may need to be further justified and investigated. This paper requires prior knowledge of the environment dynamics that must be differentiable (or partially differentiable) in nature. Many real-world reinforcement learning problems may not satisfy this requirement. Hence, the practical usefulness of the new algorithm remains a bit questionable. The authors are highly recommended to clearly evaluate and justify the practical value of the new algorithm. Furthermore, if the environment model is known in advance, it is possible to conduct effective model-based policy training without using analytical policy gradients. It remains unclear to me what the advantages of using analytic policy gradient techniques would be, compared to planning and other model-based reinforcement learning techniques. It is also desirable if the performance strength of using the analytic policy gradients, compared to other model-based reinforcement learning methods, can be experimentally evaluated and reported in this paper. The authors stated several times in the paper that the analytical policy gradients would be highly biased when they are incomplete. I don't understand what this means and why it is true. Incompleteness and bias present two separate dimensions regarding the quality of analytical policy gradients. Why should they be strongly correlated? Some mathematical claims require more clarity. For example, Lemma 4.4 mentioned a particular advantage function A-hat. However, the actual definition of this function is not presented, making it hard to accurately understand this lemma and its applicability. Some critical details were missing regarding the new algorithm design in the main text. For example, the new algorithm requires you to adjust alpha from time to time. However, it is not clear how alpha is actually adjusted. The formula for adjusting alpha is not presented and clearly justified in the main text. Meanwhile, some claims in the algorithm design also need strong justifications. For example, it is stated on page 6 that by constraining the determinant of a matrix to be near 1, we can guarantee stable policy updates. However, the validity of this statement is not proved or properly explained. Moreover, if all eigenvalues are far greater than 1, alpha would be set to be very close to 0. In this case, would the analytical gradient become useless? Similarly, it is not clear why the ratio of the two determinants gives us the difference between the two policies. According to the experiment results presented in the main text, it seems that the new algorithm is only useful when the problem being solved has many poor local optima (hence the algorithm needs to inject some uncertainties or conducts more explorations) or when the complete analytical policy gradient cannot be obtained. On problems of differentiable physics simulations, using the analytical policy gradients alone appear to be able to achieve the best results. In view of this, the real practical value of the new algorithm may need to be further investigated on more benchmark problems.

Questions

What are the possibilities and potential limitations of using the proposed policy gradient combination technique on other algorithms, particularly off-policy algorithms? What is the practical impact of the assumption regarding the prior knowledge of the environment dynamics that must be differentiable? What are the theoretical and practical advantages of the new algorithm design, compared to other model-based reinforcement learning methods? Incompleteness and bias present two separate dimensions regarding the quality of analytical policy gradients. Why should they be strongly correlated?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited 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

3 good

Presentation

3 good

Contribution

2 fair

Limitations

I do not have any concerns regarding this question.

Reviewer 39SH2023-08-11

Thank the authors for their response

I would like to thank the authors for their response, which has addressed some of my concerns. I will increase my rating a bit. In the meantime, I am not fully convinced by the practical applicability of the algorithm. Regarding the discussion on model-based RL, my point is when the model is known (hence no cost is involved in learning the model), similar to the condition of the newly proposed algorithm, how well the new algorithm can outperform model-based RL in terms of both theoretical and empirical advantages. Additionally, I still don't understand the difference between completeness and bias based on the authors' explanation. I think a more thorough mathematical definition is required. This is also the case regarding the explanation on the ratio of the two determinants and why this can measure the difference between two policies. Finally, I believe the main text of the paper should be self-contained, without relying on any appendices. Hence, the definition of A hat should be in the main text.

Authorsrebuttal2023-08-14

Thank you for raising the rating and additional comments, we appreciate them! Here are our answers to additional questions. **Q1. When the model is known, hence no cost is involved in learning the model, how well the new algorithm can outperform model-based RL in terms of both theoretical and empirical advantages.** A1. According to [1], we believe that your suggested model-based RL with known models falls into the same category as models like AlphaZero [2], which leverage planning methods such as Monte Carlo Tree Search (MCTS). To briefly recap how MCTS works: since the world dynamics model (which are the Go rules for AlphaZero) is fully known, it "simulates" future steps of gameplay using the estimations of its neural network as a prior. That is, it rolls out future gameplay steps based on rough suggestions from the neural networks, builds a search tree based on them, and selects the action that yields the best reward. In this approach, the model is fully known and is thus used as a tool to simulate various scenarios. However, we’d like to note that this strategy works in a very different environment than ours - it works in the **discrete** action space, while our method is for **continuous** one. It is unclear if we can apply planning methods like MCTS to our case or not, and thus it is difficult to compare it directly to our method. In spite of that, we’d like to point out that even if we could apply MCTS-like planning methods to our physics simulation problems, it could require much more computational cost for training. This is because it has to expand its search tree by taking a large number of timesteps, and take a SINGLE action based on it. For instance, AlphaZero runs 1600 simulation steps to build a search tree for determining a SINGLE action. Considering that the most of the computational cost comes from the expensive simulation costs for physics simulations, it is prohibitive to run such a number of simulation steps for determining action at a single time step, and thus training the network. Therefore, we’d like to underline the difficulty to directly compare model-based RL with a fully known model to our method. At the same time, the planning method like MCTS used for such model-based RL methods could require much more computational cost than ours in the worst case scenario. Exploring how analytical gradients can be efficiently applied to model-based RL methods would be another very exciting research direction. **Q2. I still don't understand the difference between completeness and bias based on the authors' explanation. I think a more thorough mathematical definition is required. This is also the case regarding the explanation on the ratio of the two determinants and why this can measure the difference between two policies.** A2. First, after some reconsideration, we realize that ‘completeness' is not an appropriate term to use in the broader, general context. We will remove the term in the revision and explain our method using bias only. However, it does not change our main claim that our method can leverage possibly biased analytical gradients better than the other baseline methods. Thank you again for pointing it out to improve the exposition of the paper. About the second issue, we’d like to reemphasize that the ratio of the two determinants does not “measure” the difference between the two policies. Instead, our view is that it is more likely that the two policies are more different from one another when the ratio deviates more from 1, at least locally, because the ratio is 1 when the two policies are the same. In the revision, empirical results will be included to support this claim. **Q3. I believe the main text of the paper should be self-contained, without relying on any appendices. Hence, the definition of A hat should be in the main text.** A3. We agree with your point, and will move the definition of A hat to the main paper. Thank you for your suggestions. [1] Moerland, Thomas M., et al. "Model-based reinforcement learning: A survey." [2] Silver, David, et al. "Mastering the game of go without human knowledge."

Reviewer 39SH2023-08-18

Thank the authors for the further clarification

I would like to thank the authors for the further clarification. Regarding model based RL, I am actually referring to existing works such as below M. Janner, J. Fu, M. Zhang, and S. Levine. When to trust your model: Model-based policy optimization. Advances in neural information processing systems, 32, 2019. Assuming that the model is known in advance without uncertainty, I think a direct comparison with such model-based policy optimization algorithms is possible.

Authorsrebuttal2023-08-18

Thank you for the reviewer's comment. First, we’d like to highlight the fact that MBPO algorithm in the suggested paper leverages (an ensemble of) neural networks to **approximate the environment dynamics**. Therefore, **it does not assume the fully known model**. In fact, if you look into the A3 of the first rebuttal above, we referred to SE-MBPO, which is an extended version of MBPO using analytical gradients. Please refer to [1] for more information. Therefore, it is not a proper method that we can use when we assume the fully known model. Even if we might be able to use the fully known model in the place of approximated model in the algorithm, then we expect that the algorithm would not be different from Soft-Actor-Critic (SAC) algorithm, which MBPO uses for backbone algorithm for policy learning. Also, we'd like to emphasize that our method does not assume we have a fully known model at hand. If we can get analytical gradients in another way, that would also work - for instance, we can approximate the world dynamics model with neural networks as MBPO does, and leverage analytical gradients coming from the approximated model. However, we believe that it is a quite different setting than ours, because now we have to consider the model approximation, which introduces a lot more things to consider in the training process. Along with leveraging Monte Carlo Tree Search (MCTS), these are all exciting future directions to further explore the trade-off between accuracy and efficiency, possibly robustness and generalization, in exploiting use of analytical gradients in model-based RL. [1] Qiao, Yi-Ling, et al. "Efficient differentiable simulation of articulated bodies."

Reviewer DV216/10 · confidence 4/52023-07-03

Summary

This paper introduces a novel policy learning method that adopts analytical gradients into the PPO framework without the necessity of estimating LR gradients. The authors introduce an adaptive α-policy that allows us to dynamically adjust the influence of analytical gradients by evaluating the variance and bias of these gradients. The experiments show that the proposed method, GI-PPO, strikes a balance between analytical gradient-based policy updates and PPO-based updates, yielding solid results compared to baseline algorithms in various environments.

Strengths

This paper introduces original ideas about how to incorporate analytical gradients into the PPO framework, taking their variance and bias into account. As a result, a locally superior policy, adaptive $\alpha$-policy is introduced and fits into PPO framework seamlessly, by dynamically adjusting the influence of analytical gradients. This paper is of solid technical quality. It provides theoretical derivations and analysis to demonstrate how $\alpha$-policy fits into PPO framework seamlessly, the relationship between reparameterization RP gradient and $\alpha$-policy, and theoretical guidance to adjust $\alpha$ based on variance and bias. The experiments are extensive in general, demonstrating the solid performance compared to baseline in different environments. Last, the paper is clearly written in general.

Weaknesses

In Section 5.1 (page 7, line 229), Figure [8] is missing in main paper, although we may find the figure in Appendix. In Fig. 3, the plot is too small to distinguish the relative performance for different algorithms. In Section 5.2 (page 8), there is typo line 259, it should be Figure [4] instead of Figure [5] for Differential Physics Simulation. In Fig. 4(b,c), GI-PPO is no better than RP for tasks Ant and Hopper in term of returning rewards. To strengthen the impact of the work, it might be better to add more challenging tasks, for example, Humanoid in Mujoco environment.

Questions

This paper is well written in general. However, for the readers to better understand the algorithm, can we move Algorithm 1 GI-PPO from Appendix and add into main paper? To utilize analytic gradients, it seems GI-PPO needs more extra computations from back-propagation. Can the author provide the computational cost comparison to different baseline algorithms?

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

There is no code provided to reproduce the results.

Reviewer 8kSu4/10 · confidence 4/52023-07-12

Summary

This paper integrates analytical gradients with PPO. They introduce an \alpha-policy to control the bias and variance during the update of the policy. Results on some differentiable environments show the effectiveness of the proposed method.

Strengths

This paper study how to incorporate analytical gradient into existing Rl algorithm, which is interesting.

Weaknesses

The main weakness of the paper is that the writing of the paper is somewhat poor. The presentation lacks motivation or necessary intuitive explanations. There are some imprecise statements and typos (see below). These flaws make the paper hard to follow. Major concerns: - This paper claims that it's a combination of analytical gradient and PPO. However, I didn't see where you use PPO (eq. 4 is not PPO!). - Some statements lack motivation. I don't understand why the algorithm is composed of three parts (Line 182). Could you give more explanation? And why the adjusting $\alpha$ should be designed in that way? Besides, it would be good if there was more explanation of Def 4.1 - The proposed method does not perform well on the physical simulation tasks. In Figure 4, the proposed method performs worse than the RP method on two high-dimensional tasks: Ant and Hopper. - In most of the practical tasks, it is hard to get the analytical gradient. This limits the potential of the proposed method. The paper only does experiments in simulation tasks. Could you give more examples of where we can get analytical gradients in practice? There are some minor technical issues in the paper: - Line 106: The symbol $A$ is repeated with the action space $A$ (in Line 97). - Line 97: You assume the probability to be within $[0,1]$. This means that the action space is discrete. However, in Line 101, you assume action space is continuous. Some statements in the paper are wrong: - "PPO relies on the observation that we can evaluate a policy πθ with our current policy πθ¯". This is not right. Because the $\rho_{\pi_\theta}(s)$ is not available. - Line 137: "the difference between πθ¯ and πθ must be small enough to get an accurate estimate." In fact, in the theory of TRPO, they introduce a lower bound with KL divergence penalty term. They don't care whether that's an accurate estimate.

Questions

- What do you mean by inherent chaotic nature (Line 25)? Could you please give some examples? - I'm confused about how is eq. 2 used in your method. Please also see my comments in the weakness.

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

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

2 fair

Presentation

1 poor

Contribution

3 good

Authorsrebuttal2023-08-16

Dear Reviewer 8kSu, Did our rebuttal address issues you raised? Do you still have any more questions? Thank you. Best, The Authors

Reviewer WNVi7/10 · confidence 4/52023-07-20

Summary

In this manuscript, the author(s) combined the PPO framework with analytical gradients and proposed a policy learning method to learn better policy quickly. Through empirical experiments, the author(s) demonstrated that the proposed method was superior to baseline methods in several scenarios.

Strengths

### Strong points: 1) Utilizing $\alpha$-policy to deal with the influence of analytical policy gradients when training; 2) Adopting criteria to detect/calculate variance and bias; 3) An algorithm was designed to handle the strength of analytical gradients during updating; 4) Empirical results look good. By the way, in the Supplementary Materials, the author(s) provided a mp4 video file, where the animation on the scenario of the traffic problem looks fantastic. Thanks for providing that.

Weaknesses

### Weak points: Although the work of this manuscript looks interesting, some points in this manuscript are not so clear to me, I hope the author(s) can explain/detail them. Thanks. 1) Some descriptions of empirical experiments are not so clear. 2) Missing codes related to the proposed method in this manuscript. 3) If possible, some theoretical results in this manuscript should be empirically verified.

Questions

Detailedly, I have the following comments/questions: 1) For the description or empirical experiment details, some places in this manuscript seem unclear. For example, (1) Definition 5 was mentioned several times in the manuscript, such as, in Lines 179, 188, and even in the Supplementary Materials, Line 419, but where is Definition 5? (2) The proof is mainly for Proposition, Lemma, Theorem, etc. In Line 419 of Supplementary Materials, it mentioned "Proof of Definition 5", I am not clear. (3) Could you please explain/clarify the fluctuation of the blue curve (i.e., GI-PPO) in Figure 3 (c)? Why caused this phenomenon? (4) The order number (or citation) of Figures in this manuscript is messy. For instance, in Line 229, Figure 8 is in the Supplementary Materials, right? Figure 1 seems to be cited until Line 245? 2) If possible, some theoretical results, for example, Proposition 4.5, can be verified by empirical experiments? 3) Some other tiny issues/typos: (1) The font size used for the coordinates in Figure is too small. (2) The format of mathematical formulas in this manuscript is inconsistent. Some end with punctuation, and some don't. (3) Many issues in References. For instance, Sometimes the conference name is the full name plus the abbreviation, and sometimes only the full name; Missed source of the cited references; and so on. Hope the author(s) will check carefully and correct them. In addition, I did not scrutiny the proofs step-by-step, but I think the proofs should be Ok. Thank the author(s) for submitting this interesting work. Except for the several possible issues/questions mentioned above. I think this is a basically good manuscript. I look forward to the response from the author(s). If the author feels any of my comments are inappropriate, please feel free to point them out :-) Thanks.

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

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

4 excellent

Contribution

3 good

Limitations

Yes. As the author(s) revealed, there might be several limitations, such as(1) How to approximate $\alpha$-policy efficiently; (2) Adjusting PPO's clipping range; (3) Computational efficiency. Hope that the author(s) can further investigate these limitations and solve these potential problems in the near future. Good luck!

Reviewer xDc87/10 · confidence 4/52023-07-27

Summary

This paper introduces a policy gradient algorithm based on Proximal Policy Optimization (PPO) algorithm. This new algorithm utilizes the analytical gradients from differential environments and achieves competitive performance in various scenarios including function optimization, classic control environments, and traffic control environments. To integrate the analytical gradient into the PPO framework, they use the reparameterization trick and introduce the concept of $\alpha$-policy which is a locally superior policy. When $\alpha$ is sufficiently small, the difference between the original policy and this $\alpha$-policy would be small enough to satisfy the difference constraint in the PPO algorithm. Moreover, they propose metrics considering variance, bias, and out-of-range-ratio to dynamically adjust $\alpha$ during the training.

Strengths

1. The idea of utilizing analytical gradients in physical control environments is powerful and can dramatically increase the efficiency of policy gradient algorithms when analytical gradients are available. This paper advances this idea by integrating it with the PPO algorithm and achieves competitive performance in various environments. 2. This paper proposes the concept of $\alpha$-policy and connect it with the reparameterization trick. $\alpha$-policy is an adjustable policy such that when $\alpha$ is small enough the difference between $\alpha$-policy and the origin policy can be small enough to fit into PPO algorithm. 3. This paper proposed metrics considering variance, bias, and out-of-range-ratio to dynamically adjust $\alpha$ during training.

Weaknesses

1. When trying to fit the analytical gradients into the PPO algorithm, this paper didn't consider dynamically changing the PPO clip limit. Thus, the analytical gradients are not fully utilized. 2. More theoretical analysis and better ways on how to adjust $\alpha$ during training could be done and proposed. 3. The experimental performance showed improved performance compared to PPO but does not achieve state-of-art performance in some environments.

Questions

1. Line 419 "Proof of Definition 5" does not match the actual definition index. The author should separate definition 4.1 into a definition and a proposition. 2. Can the author integrate the variance reduction techniques in [Xu et al., 2022] to achieve better performance? If not, what changes can be made to accommodate this?

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

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

I did not identify any potential negative societal impact from this work.

Reviewer WNVi2023-08-12

Thank the author(s) for the responses

Thank the author(s) for the responses. I read the reviewers' comments and the rebuttals from the author(s). The author(s) answered my questions and cleared up my doubts. I increase my score and vote to accept this manuscript, but on the condition that the author(s) will carefully make the changes and correct the issues in the final version, as claimed in responses/rebuttals. Thanks and good luck!

Authorsrebuttal2023-08-14

Thank the reviewer, we really appreciate the support! We’ll carefully revise the paper, as the reviewers suggested. Thanks.

Reviewer DV212023-08-15

Thanks to the authors for the rebuttal

I’ve read comments from all the other reviewers. Thank you for your rebuttal, and I appreciate that my concerns have been addressed.

Reviewer xDc82023-08-17

Thank the authors for the responses and supplementary materials

I would like to thank the authors for the responses and supplementary materials. They have addressed my questions. I encourage the authors to integrate discussions in supplementary materials to the main paper because they provide more direct insight on $\alpha$.

Authorsrebuttal2023-08-17

Thanks for the support, we will revise the paper as the reviewers' suggestions, thank you.

Area Chair Ezmh2023-08-19

To Authors: Thank you for submitting your rebuttal. We appreciate your efforts in addressing the reviewers' comments and providing clarification. To Reviewers: We kindly request you to complete your response to the authors' rebuttal as soon as possible. Time is of the essence, and the deadline is fast approaching. Your timely feedback and expertise are crucial in ensuring a fair and thorough evaluation process. Please prioritize reviewing the authors' response and provide your final feedback accordingly.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC