Adjustable Robust Reinforcement Learning for Online 3D Bin Packing

Designing effective policies for the online 3D bin packing problem (3D-BPP) has been a long-standing challenge, primarily due to the unpredictable nature of incoming box sequences and stringent physical constraints. While current deep reinforcement learning (DRL) methods for online 3D-BPP have shown promising results in optimizing average performance over an underlying box sequence distribution, they often fail in real-world settings where some worst-case scenarios can materialize. Standard robust DRL algorithms tend to overly prioritize optimizing the worst-case performance at the expense of performance under normal problem instance distribution. To address these issues, we first introduce a permutation-based attacker to investigate the practical robustness of both DRL-based and heuristic methods proposed for solving online 3D-BPP. Then, we propose an adjustable robust reinforcement learning (AR2L) framework that allows efficient adjustment of robustness weights to achieve the desired balance of the policy's performance in average and worst-case environments. Specifically, we formulate the objective function as a weighted sum of expected and worst-case returns, and derive the lower performance bound by relating to the return under a mixture dynamics. To realize this lower bound, we adopt an iterative procedure that searches for the associated mixture dynamics and improves the corresponding policy. We integrate this procedure into two popular robust adversarial algorithms to develop the exact and approximate AR2L algorithms. Experiments demonstrate that AR2L is versatile in the sense that it improves policy robustness while maintaining an acceptable level of performance for the nominal case.

Paper

References (60)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer Ua6d4/10 · confidence 4/52023-07-05

Summary

This work proposed a novel adjustable and robust Reinforcement Learning framework for Online 3D Bin Packing task. The proposed method can achieve balance between performance and the worst-case environment.

Strengths

The writing is clear, and the experimental results demonstrate that the proposed method works well compared with other RL methods for the targeted 3D Bin Packing problem.

Weaknesses

- In the section of “Training and Evaluation Setting”, the author mentioned that they are using two different settings which are discrete and continuous. However, I did not find the table showing the experimental results using continuous setting. - The author mentioned that they use three metrics(U_ti, Std and Num) to evaluate the performance. But only the best U_ti are highlighted in Table 2. I do think the analysis for Table 2 needs more details especially for the Std and Num. - The computation complexity is also crucial for the RL algorithms. The comparison of computation complexity should be analyzed.

Questions

- Table 2 is the main result, where larger N_B and larger beta represent the harder problem, and we can see that throughout the table, the baseline RARL often achieves the best results, and the proposed Exact AR2L and Approx AR2L is inferior. What is your interpretation? - How to ensure the Approx AR2L can approximate the AR2L in the "real-world" setting? Since you do not perform experiments in the real world? I mean, can your added randomness correspond to a real-world robot arm use case (in some industry), and if so, what kind of distribution can characterize such randomness? I believe this must be addressed by data-driven experiments; however I do not see this in your paper. - Attacker: I understand the "attacker" is used to generate proper samples to train your DRL algorithm for policy improvement. However, do you have a better name for such adversarial or adversary other than an attacker, as they are used in a normal setting. If there is a true attacker of your DRL training, they will probably attack in other places the let the framework down. I understand this might be due to the term already used in the literature, I just want to not confusing the reader about technical terms.

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

3 good

Contribution

3 good

Limitations

There is no such limitation in this aspect for this paper.

Reviewer HsT15/10 · confidence 4/52023-07-06

Summary

For solving the online 3D Bin Packing Problem (BPP), the authors employ an iterative procedure to search for relevant hybrid dynamics and refine their corresponding strategies. By optimizing the weighted sum of returns, AR2L algorithm which improves the robustness of the packing policy achieves a balance between nominal insurance performance and worst case insurance performance.

Strengths

In experiment, the authors evaluate the robustness of six heuristic method. Moreover, this paper gives the exact AR2L algorithm and Approximate AR2L algorithm which has positive results. This two part of solving online 3D Bin Packing Problem has been proved to have relatively good results theoretically and experimentally.

Weaknesses

Experimentation is not sufficient. Continuously valued perturbations for addition may not correspond to distributions in the physical world. Is there any difference between the response of the baseline to the disturbance of these continuous values and the method in the paper to the disturbance of these continuous values. Whether there is a simulation or a robotic arm, a demo for dealing with extreme situations.

Questions

1. For Figure 2, it should be ensured that the positions of the same modules remain unchanged, and the comparison changes should be highlighted, so as to achieve of the exact and approximate AR2L algorithm 2. The relationship between offline and online BPP introduced by introduce should be clarified, if not needed, offline should be deleted. Because the whole article does not focus on online BPP 3. At the same time, please supplement the experimental part in weaknesses

Rating

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

3 good

Presentation

2 fair

Contribution

3 good

Limitations

due to the mixture-dynamics attacker, it will introduce additional calculations to increase complexity. At the same time for the latency of the whole process will be increased as well.

Reviewer HCvp7/10 · confidence 5/52023-07-06

Summary

This paper investigates the online three-dimensional bin packing problem (3D-BPP) and extends the PCT algorithm by proposing an adjustable robust reinforcement learning (AR2L) framework that balances the performance of policies in average and worst-case scenarios. The paper designs a permutation-based adversary and introduces an objective function that combines expected return and worst-case return with weights. A lower bound is derived to guide policy learning. The paper presents two algorithms to implement the AR2L framework, one exact that requires training a hybrid dynamics adversary, and another approximate that samples from the original dynamics and permutation-based adversary. Experiments are conducted in both discrete and continuous settings, demonstrating the effectiveness and superiority of the AR2L framework.

Strengths

1. The problem addressed in the paper is highly significant. The authors extend the state-of-the-art PCT algorithm by optimizing its worst-case performance through an adversarial approach. I also have a strong impression of the PCT work, as it intuitively and effectively solves the important online bin packing problem. It is great to see the functionality of PCT being further enhanced through this work. 2. I appreciate the organization of the paper. The authors demonstrate a strong understanding of packing and reinforcement learning, supporting each argument with data or theory. The paper is clear, understandable, and reasonable, making it enjoyable to read. 3. The proposed approach is clever. The authors introduce a novel robust reinforcement learning framework, where a permutation-based adversary is designed to generate worst-case problem instances. This is an innovative and practical method. Additionally, the paper derives a lower bound for an objective function and utilizes it to guide policy learning, providing a theoretical contribution. 4. The paper presents two algorithms to implement the AR2L framework, one exact and one approximate, catering to different scenarios and requirements. 5. The paper conducts extensive experiments, comparing the AR2L framework with multiple packing and benchmark methods, demonstrating its advantages in terms of average performance and robustness.

Weaknesses

The paper does not provide guidance or suggestions on how to choose the hyperparameters α and ρ. I noticed that the optimal values of α vary when β changes in the best-performing AR2L algorithm. Perhaps this could be considered as future work.

Questions

1. If I understand correctly, in Table 2 of the main text and Table 2 in the appendix, each trained policy is still tested using data generated by the attacker (as the performance of the baseline algorithms varies with different beta values, indicating a change in the tested data). So, if we remove the attacker and everyone uses the same data for testing, what would be the performance of AR2L? I am very interested in the practical implications of packing policies and would like to know if the packing policy learned through adversarial training still has advantages when tested on normal data (e.g., object sizes following a uniform distribution). This inquiry is purely out of curiosity and will not affect my scoring. 2. In the appendix, the authors mention that they failed to reproduce some PCT results. Based on my own attempt, I believe PCT is quite reliable. I suggest that the authors contact the PCT authors as they may be able to assist in successfully reproducing the results. It is possible that there are some issues related to the usage or implementation, as packing involves various complex settings that require consideration of multiple aspects, such as discrete domain, continuous domain, stability checks, and data types. The functionality of the code can be quite complex. I believe the authors of PCT will be willing to help you successfully reproduce the results, and I have sent them a message to notify them and draw their attention to possible coming inquiry emails. 3. A quick question: I would like to reproduce the results mentioned in the paper. In the code provided, what does "bal_policy" represent? Does the code refer to the code for ExactAR2L or ApproxAR2L, or is there an option to switch between the two algorithms?

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

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The paper only considers one type of permutation-based attacker and does not explore other possible attack methods, such as adding or removing items, modifying the size or shape of items, etc. ----------------------After Rebuttal---------------------- The reviewer appreciates the author's response. However, the reviewer is still considering whether it is necessary to propose a new candidate generation method. During the rebuttal period, the reviewer downloaded the official implementation of PCT. The final test results were even higher than those reported in the PCT paper. Because the reviewer does not see a direct response in the rebuttal, holding onto this doubt, the reviewer intends to slightly lower the rating.

Authorsrebuttal2023-08-17

We appreciate the feedback provided in the review. In fact, we found that the official implementation of the baseline method PCT was just updated 3 weeks ago. Consequently, we were only able to download the older version of PCT before the Neurips deadline. When attempting to reproduce PCT following the official instructions, we did not achieve satisfactory performance in the continuous setting. As a result, we made the decision to use a new heuristic method for candidate generation in our experiments. In addition, we would like to clarify that our study specifically focuses on the robustness of the packing policy, rather than exploring alternative candidate generation methods. Therefore, to maintain a fair comparison in the continuous setting, we ensured that the same heuristic method was used for generating all candidate positions across all approaches. We appreciate the reviewer’s thorough understanding and insightful comments regarding our method. We are grateful for the valuable feedback provided, and will emphasize in the revised paper about the settings of the candidate generation method. If there are any additional concerns or questions, we are more than happy to provide further explanations and address them to the best of our abilities.

Reviewer HCvp2023-08-18

Thank you for your response. As I am interested in the packing problem, I hope to avoid unnecessary complexities that might confuse the community. I took the initiative to reach out to the authors of the PCT paper in hopes that they could provide assistance and update their usage documentation. However, in the authors' response to this article, I did not observe them presenting any persuasive experimental results, which is why I lowered my rating. Nonetheless, because I still find great value in this paper, ensuring the worst-case performance for the packing problem is meaningful. I acknowledge the authors' efforts; however, I urge them to thoroughly replicate the performance of existing methods in the revision. Based on these conditions, I am inclined to raise my recommendation to a score of 7.

Authorsrebuttal2023-08-18

We sincerely appreciate the reviewer’s insightful comments and valuable feedback on our work. Your concerns and suggestions are highly valued and contribute to the packing problem research community. We acknowledge the importance of thoroughly reproducing the performance of the baseline methods using the latest official implementation, as you have recommended. In the revision, we will carefully address this concern and ensure accurate and up-to-date comparisons with the baseline methods. We are grateful for your valuable feedback, which has helped us improve the quality of our work.

Reviewer Em3c6/10 · confidence 2/52023-07-06

Summary

This work addresses the problem of 3D bin packing problem (3D-BPP). Specifically, it develops a permutation-based attacker and subsequently proposes an adjustable robust reinforcement learning (AR2L) framework. This allows an algorithm to consider both the average and worst-case performance with the attacker, where the packing objective is a weighted sum of expected and worst-case returns over the space utilization rate. The proposed framework is integrated with prior work to develop the exact and approximate variant of the proposed AR2L. -------------------------- I acknowledge the author's effort in the rebuttal and have made changes to the review accordingly.

Strengths

+ This work proposes to train a novel permutation-based attacker (RL-based policy) that can re-order the sequence of the observed items to reduce the space utilization of bin packaging tasks. The permutation-based attacker is used to quantitatively evaluate a given algorithm's performance under the worst-case scenario. The empirical results show new insight into the robustness of existing works. + This work developed an adjustable robust reinforcement learning (AR2L) approach, which learns a packing policy based on both average scenarios and worst-case scenarios. This can be achieved via a surrogate problem where an optimal policy can be identified with the maximal lower bound. + Two forms of implementation of the AR2L algorithm are shown in this work. The exact AR2L algorithm performs better worst-case performance under the worst-case attack with additional computation to the entire framework. On the other hand, the approximate AR2L algorithm introduces estimation errors with the cost of performance drop. For both variants, it outperforms the corresponding baseline methods.

Weaknesses

- This work shows comprehensive evaluation with different combination of \alpa, \beta, and N_B parameters. While this provides valuable insights, it is unclear what would be the most optimal configuration in real-world deployment since the conditions could vary across task.

Questions

- In eqn 2, should the last term be \alpha d( P^m || P^w )?

Rating

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

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central 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

3 good

Limitations

No immediate limitations or impact raised from the novelties from this work.

Reviewer fVdw4/10 · confidence 4/52023-07-25

Summary

The paper is in the category of work that uses reinforcement learning for combinatorial optimization problems. More specifically, it focuses on the online version of the 3d-bin-packing problem (3D-BPP), and proposes a robust RL solution to address the uncertainty that arises from the permutation of an adversarial nature. At the core of their solution, Adjustable Robust Reinforcement Learning (AR2L), is 1) a new attacker that is based on permutation of the observable sequence of the items, 2) a new objective function that adjusts weights between nominal performance and worst-case performance, and 3) theories and algorithms that solve the new optimization.

Strengths

+ RL for combinatorial optimization is an interesting and trending topic. It can potentially draw interest from the RL/OR community. + Robustness in 3D-BPP is a practical and important issue. + Permutation-based attacker makes sense. + The adjustable objective makes sense. + The evaluation of robustness for existing methods is applausible. + Connections of AR2L to RARA and RfMDP are interesting.

Weaknesses

- The attacker capacity seems questionable. Because of the following reasons, the problem setting seems rather impractical. These make the formulation, theorems, and evaluation somewhat trivial. 1) If you formulate the nature as an adversary, then the nature should have control over the permutation of the entire sequence of items that are not presented to the packing policy, not just the observable ones. 2) Also, it seems that the evaluation assumes the packing policy only observes one item. This is a bit weired because in this case the attacker does not have any capacity. 3) Moreover, the attacker's policy is to change only one item to the most preceding item, but the attacker should be able to change the entire permutation. This is really how the attacker makes the training robust, and evaluations should be conducted against such an attacker. 4) Finally, the setting that the sequence of the observable items can be changed is also weired. This is because the packing agent/policy has already seen the set of observable items, you cannot suddenly change them without the notice from the packing agent. - The paper, along with its formulation, seems to be too much restricted to robustness to permutation-based attack in 3D-BPP. Do the theories/results generalize to other settings? There is a lack of discussion from that aspect. This makes the technical contribution rather limited. The paper would be stronger if it can shed light on how the algorithms are generally applicable to more combinatorial optimization problems with uncertainty. - Theorem 1: $\gamma$ is not explained -- is it the discount factor? Isn't is assumed to be 1 in the Preliminaries section? If so, the denominator is 0. At least $1-\gamma$ is a very small value, making the second term on the RHS of Equation (2) dominating the entire RHS. This significantly restricts the meaningfulness of Theorem 1 (a major theoretical contribution). It seems that the objective in Equation (3) makes sense by itself, but without Theorem 1 this is only a heuristic. Also, why the coefficients are thrown out in Equation (2) and turned into Equation (3) was not explained. - The empirical results are not convincing 1) it looks like that the average performance of AR2L is outperformed by the baselines (esp. RARL) in many cases (esp. in the most challenging setting where $\beta=100$. It is surprising that PCT is not the best even for the nominal setting ($\beta=0$). These have not been discussed and explained. 2) Also, although the paper is about robustness, it primarily focuses only on average performance (Uti), discussions of Std are limited even thought the results are shown. 3) I was expecting to see some guidelines/lessons-learned from what values I should take for the hyperparameters $\alpha$ and $\rho$, but did not see. If I were to use the algorithms, how do I choose them? It seems that if I choose the wrong ones, I might end up with a solution that is worse than the baselines. The mysterious performances regarding hyperparas are very un-convincing. Minor issues: * Refereces to equations should be Equation (1), instead of 1. See e.g., Line 205. * A typo at the end of Equation (2). $P^0$ -> $P^w$. *

Questions

See questions in the weakness part.

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

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

NA

Reviewer Em3c2023-08-16

I acknowledge that I have read the rebuttal and the authors has responded to my concerns.

Reviewer fVdw2023-08-16

I thank thre authors for the responses. However, I am not entirely convinced by the issues around the attacker's capacity, the empirical evaluation, and the hyperparameters. In terms of the attacker's capacity, e.g., why doesn't the attacker permute the entire sequence of observable items instead of just putting one item ahead of the others? Although eventually the attacker will change the order of the entire sequence via multiple steps, this limits the power of the attacker. Also, why doesn't the attacker have the capacity of changing the item distribution (like Kong et al.)? In terms of the empirical evaluation. E.g., it is not convincing why PCT is outperformed by AR2L even when $\beta=0$. Does this imply that the baselines are weak in the first place? Also, the std of AR2L(1) is larger than RARL in many cases -- this is counter-intuitive since RARL focuses on the worst-case, it should have higher stds. In terms of the hyperparas, how about $\alpha>1$? $\alpha=1$ is the largest value being tested, but it seems that larger $\alpha$'s may yield even better performances. But this may indicate that the "robustness" does not matter that much and the algorithm reduces to maximizing the nomimal objective.

Authorsrebuttal2023-08-17

Thank you for your response. We sincerely appreciate your valuable feedback. We would like to provide further clarification and address these concerns. Q1: The settings of attacker **Moving one item** From the perspective of the practical setting of online BPP, the packing policy can observe multiple items, but it can only pack the first item into the container. Let's consider a scenario where the attacker permutes the entire sequence of observable items at timestep $t$. In this case, regardless of the permutation applied to the remaining items at timestep $t$, at timestep $t+1$, the remaining items will combine with a new incoming item to form a new sequence that will be further permuted. Therefore, any permutation of the remaining items at timestep $t$ will be ignored and subsequently permuted with the new incoming items. From the perspective of the influence on the packing policy, let's consider a simplified value function $V(C_{t}, B_{t}) = r_t + V(C_{t+1}, B_{t+1})$, where $C_{t}$ and $B_{t}$ represent the bin configuration and the permuted observable item sequence, and $r_t$ is the reward. It tells us that the value of the packing policy at timestep $t$ depends on both the reward $r_t$ and the value function $V(C_{t+1}, B_{t+1})$. And $r_t$ and $C_{t+1}$ are influenced by the first item in $B_{t}$. The remaining items in $B_{t}$ will combine with a new item to form a new sequence, which will undergo further permutations as $B_{t+1}$. Consequently, the permutation of the remaining items at timestep $t$ will be disregarded in the attack process. From the perspective of the implementation, if we allow the attacker to permute the entire sequence of observable items, the action space will grow exponentially as the number of observable items increases. This large search space may have a detrimental effect on the convergence of the RL-based attacker. **Item distribution** In our paper, we specifically focus on the impact of item permutation on the packing policy, rather than the item distribution. To ensure that our study remains focused on the randomness introduced by item permutation, we design the permutation-based attacker, which allows us to avoid any influence stemming from changes in the item distribution. Furthermore, as mentioned in PCT, larger items can simplify the scenario, while smaller items can trickily make the task more challenging. Therefore, we aim to prevent the attacker from simply learning to selectively favor certain types of items to create challenging instances. Instead, we want the attacker to genuinely learn how to permute the item sequence in order to identify and create challenging scenarios. Hence, we have designed an attacker that is unable to alter the item distribution in order to maintain the integrity of our study. Q2: The evaluation **The performance of $\beta=0$** We would like to emphasize that there are inherently challenging instances present in the nominal dynamics ($\beta=0$). Therefore, one approach to improving performance when $\beta=0$ is to enhance the performance specifically on those challenging instances while maintaining performance on others. When using a smaller value of $N_B$, the distribution of training data (mixture instances) and testing data ($\beta=0$) exhibits less deviation. This enables the policy to effectively handle both challenging instances and moderate instances, thereby improving generalization. However, when $N_B=20$, the substantial deviation between the distributions can negatively impact performance when testing on $\beta=0$. And we have confidence in the strength of the baseline methods chosen due to the fair experimental comparison settings. **Standard Deviation** We would like to clarify that in a high uncertainty environment, such as online 3D BPP, robustness specifically refers to the ability of a policy to consistently perform well despite the presence of uncertainties. It is intuitive that the RARL policy will consistently exhibit a conservative behavior, regardless of the problem instances provided. This conservative behavior ensures a consistent performance and reduces the variance of the policy. On the other hand, the AR2L(1) policy demonstrates a less conservative behavior, resulting in relatively larger variance. Q3: The hyperparameter We have the flexibility to set $\alpha > 1$. However, based on observations from our empirical study, we found that larger $\alpha$ can actually lead to a degradation in performance in the nominal dynamics. This suggests that AR2L with larger $\alpha$ tends to overprioritize worst-case scenarios, similar to previous robust methods. Given that our objective is to achieve a desired balance between performance in the nominal and worst-case dynamics, it is not recommended to set $\alpha > 1$. Doing so may result in an excessive focus on worst-case performance, compromising the overall performance in the nominal dynamics.

Reviewer fVdw2023-08-18

I aprpeciate the authors' clarification. I am now more convinced by the attacker setting. But I am still not convinced by the evaluation. - The performance when $\beta=0$. Intuitively, isn't a regular RL policy, or AR2L(0) the best option? - Standard deviation -- the problem is that the performances are inconsistent here. It looks like sometimes the std of RARL is smaller than AR2L(1), but sometimes it is larger (e.g., when $N_B=10, \beta=0$ or $100$). - From the table, it does not seem so. It seems that there is a trend of improving for the average performance when you increase $\alpha$. Also, this is somewhat counter-intuitive -- why $\alpha=1$ yields better average performance than, e.g., when $\alpha=0.5$ or $\alpha=0$? Isn't smaller $\alpha$'s more focused on average performance?

Authorsrebuttal2023-08-19

We express our sincere gratitude for the valuable comments provided on our work, and we are genuinely appreciative of your understanding and acceptance of our attacker settings. We are more than willing to provide further clarification on the evaluation part to address the concerns you raised. Q1: The performance when $\beta=0$. Intuitively, when testing with $\beta=0$, the regular RL policy appears to be a better option compared to AR2L with a small value of $\alpha$ (e.g., AR2L(0.3)). However, upon analyzing the empirical results, we observed that by incorporating a small quantity of challenging problem instances into the training setting of the nominal dynamics, the overall performance can actually be improved. This finding suggests that when we introduce a few challenging instances during the training process, the performance on moderate instances from the nominal dynamics ($\beta=0$) remains unaffected. This prevents an overemphasis on worst-case performance while allowing the packing policy to learn from more challenging instances. Consequently, the policy's ability to handle these inherently difficult instances from the nominal dynamics can be enhanced. We consider this observation to be an interesting finding, as it highlights the potential benefits of including a small number of challenging instances during training, leading to improved generalization. We believe that this finding provides valuable insights to the field of the packing problem and sheds light on the effectiveness of incorporating challenging instances in the training process for more general combinatorial and sequential problems. Q2: Standard deviation. Based on the observations from the empirical studies presented in Table 2 of the main text, we found that in 17 out of 20 tasks, AR2L(1) exhibits a larger standard deviation compared to RARL. Conversely, in the remaining 3 tasks, AR2L(1) shows a smaller standard deviation. We believe these results demonstrate a consistent trend. In the majority of tasks, AR2L(1) exhibits a larger standard deviation, indicating its ability to consider instances from both the nominal and worst-case dynamics. It is worth noting that even in previous robust methods, such as CPPO and RARL, their standard deviations sometimes exceed some of their respective baselines, while in other cases, their standard deviations are smaller. We believe that the consistent trend observed in the majority of tasks provides an explanation that is in line with our expectations. And we believe AR2L's formulation to efficiently adjust robustness weights is valuable compared to the relatively conservative RARL setting. Q3. To provide a thorough explanation for the observations in Table 2, let us start with the scenario where $N_B=5$. In this scenario, the attacker's ability is limited to permuting only 5 observable items. As a result, the distribution deviation between the training data (mixture data) and the testing data ($\beta=0$) is relatively small. This means that it is acceptable to increase the value of $\alpha$ and introduce more challenging instances into the training settings of the nominal dynamics. By doing so, the generalization of the model can be improved while avoiding an excessive focus on worst-case performance. However, as the number of observable items ($N_B$) increases, the distribution deviation between the training data (mixture data) and the testing data ($\beta=0$) also increases due to the attacker's improved capability. In such cases, it becomes less desirable to significantly increase the value of $\alpha$, as doing so would introduce a large number of challenging instances into the training set. This can lead to a substantial deviation between the training set and the testing set. As a result, in the cases of $N_B=10$ and $N_B=15$ under $\beta=0$, AR2L tends to favor $\alpha=0.5$. Furthermore, in the scenario where $N_B=20$, AR2L cannot outperform PCT due to the larger distribution deviation caused by the increased number of observable items of the attacker. In summary, when the attacker's capability is limited, a higher value of $\alpha$ can be beneficial. However, as the attacker's capability increases, a lower value of $\alpha$ is more suitable. Hope this address the reviewer's comment on selection of $\alpha$. Thank you for your response again. We genuinely appreciate your valuable feedback and the opportunity to provide further clarification on any concerns you may have. Your feedback is crucial to improving the quality and comprehensibility of our research. If there are any additional concerns or questions, we are more than happy to provide further explanations and address them to the best of our abilities.

Reviewer fVdw2023-08-21

Thanks for your response. Unfortunately I was not entirely convinced about the empirical evaluations. It appears to me that there is a lack of comprehensive understanding of what results were obtained and why they were obtained -- the authors seem to have a better understanding with the push of questions, but it is questionable that within a short time that these phenomena can be fully understood and well explained. But like I said, I am more convinced by the attacker model after the authors' responses. So I am willing to increase my rating from 3 to 4.

Authorsrebuttal2023-08-21

We sincerely appreciate your comments and feedback on our work. In the revised version, we will focus on discussing the empirical results in more detail to demonstrate the significance and rationality of our findings. We are grateful for your contribution in the rebuttal process, as it has helped us improve the quality of our paper.

Reviewer Ua6d2023-08-17

Comments after rebuttal

I acknowledges that I have read the author's rebuttal. I believe it has answered my questions. I think overall this is a paper with sufficient amount of efforts.

Authorsrebuttal2023-08-18

We thank the reviewer again for taking a careful review of our work, and we will incorporate the suggestions into the revised paper. We would appreciate that if the reviewer could re-evaluate the review score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC