No Regrets: Investigating and Improving Regret Approximations for Curriculum Discovery

What data or environments to use for training to improve downstream performance is a longstanding and very topical question in reinforcement learning. In particular, Unsupervised Environment Design (UED) methods have gained recent attention as their adaptive curricula promise to enable agents to be robust to in- and out-of-distribution tasks. This work investigates how existing UED methods select training environments, focusing on task prioritisation metrics. Surprisingly, despite methods aiming to maximise regret in theory, the practical approximations do not correlate with regret but with success rate. As a result, a significant portion of an agent's experience comes from environments it has already mastered, offering little to no contribution toward enhancing its abilities. Put differently, current methods fail to predict intuitive measures of ``learnability.'' Specifically, they are unable to consistently identify those scenarios that the agent can sometimes solve, but not always. Based on our analysis, we develop a method that directly trains on scenarios with high learnability. This simple and intuitive approach outperforms existing UED methods in several binary-outcome environments, including the standard domain of Minigrid and a novel setting closely inspired by a real-world robotics problem. We further introduce a new adversarial evaluation procedure for directly measuring robustness, closely mirroring the conditional value at risk (CVaR). We open-source all our code and present visualisations of final policies here: https://github.com/amacrutherford/sampling-for-learnability.

Paper

References (64)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer ycoo5/10 · confidence 4/52024-07-03

Summary

The paper investigates the limitations of popular UED approaches, such as PLR and ACCEL, demonstrating that they do not improve upon the Domain Randomisation baselines, where levels are randomly sampled. The author's main claim is that learnability does not correlate with the scoring functions MaxMC and PVL used by current UED approaches, leading to sub-optimal performance. To address this issue, they propose Sampling For Learnability (SFL), a simple algorithm where levels are chosen to prioritize learnability. In the context of the paper, the learnability of a level is defined as $p \cdot (1-p)$, where p is the success rate of the agent on that given level. The authors conduct experiments on three different environments: MiniGrid, Single, and Multi-Agent JaxNav. As a metric, they propose the use of the conditional value at risk (CVaR), evaluating on an α% of levels on which the agent performs worst, from a randomly generated buffer of levels. The results show that SFL outperforms current UED approaches. Simultaneously, when it is possible to compute it, true regret outperforms SFL, indicating that the flaw of UED algorithms is indeed in the scoring function approximation. To summarize, the main contributions of the authors are: - Demonstrate sub-optimality of current UED approaches, due to lack of correlation between their scoring function and learnability - Propose a new algorithm, SFL, which prioritizes levels with high learnability

Strengths

- UED has indeed been seen to perform sub-optimally in practice, i.e. as evidenced in the JaxUED paper. It is thus an important problem to address since limited work has been done - The investigation of the correlation between learnability and the scoring function is a good justification for the paper itself to exists, and gives a plausible reason for the poor performances of UED on some tasks. - The SFL algorithm proposed is novel (as far as I know), simple, and easy to understand. The results obtained on the 3 environments they considered seem promising. - CVaR seems a more principled and justified way to investigate the robustness of Auto-curricula algorithms

Weaknesses

- Learnability definition: why are environments that are solved exactly half of the time the most valuable ( given the $p(1-p)$ function used to define it)? It seems an arbitrary choice, which should need to be justified better through additional experiments. I do not agree that levels that are solved 5% of the time are as valuable as ones solved 95% for example. - Many parameter choices are just reported with no justification supporting it ( e.g. hyperparameter $\rho$ for the algorithm) - Experiments are poor in broadness. Given the purely empirical nature of the paper, I would expect way more experiments to indeed show that SFL outperforms the UED baselines. This includes more seed for the Multi-agent JaxNav, and more environments. As of now, it is not possible to actually confirm that SFL is indeed better than UED algorithms. - The writing is poor and sometimes confusing. Learnability is referenced way before being formally defined, the syntax of some sentences is poor, etc... By reading the paper, it seems the writing has been rushed.

Questions

- In Algorithm 1, $\rho$ is set to $0.5$. Hence, half of the levels used for updating the policy $\pi_\phi$ are ones with high learnability, and half are sampled at random. Why do not set $\rho$ to a higher value? I would like to see a plot showing the performance of SFL upon varying $\rho$, so as to have a better idea of how indeed training in 'more learnable' environments correlates with better performances - In Alg.1, what are the performances if instead of sampling uniformly from the high-learnable levels $\mathcal{D}$, you select levels in decreasing order of learnability? - What is the average/median learnability of the randomly created levels in Alg 2? What about the top-K ones? It is important to see and report it so indeed one is sure that the levels you are using are indeed ones with high-learnability. As far as I can infer from the paper, it could still be that most of the randomly created levels have low learnability. - Can you provide more details on what you did to make current UED methods work better? The poor performance of UED algorithms was already highlighted in the JAXUED paper, but I am not sure if this is just because they were not optimized fully, e.g. via a more extensive hyper-parameter search - Did you try using different definitions of Learnability? How did they perform?

Rating

5

Confidence

4

Soundness

2

Presentation

1

Contribution

3

Limitations

No ethical limitations

Reviewer ycoo2024-08-12

Response to authors

Dear Authors, Thanks for your detailed response and additional experiments. **Different values of the sampling ratio**: Thanks for producing these results. In the final manuscript, I would be interested in seeing other possible features of the environment which are common in levels sampled with SFL **Average vs top-k learnability in UED**: This result is interesting. I would consider to appropriately insert it in the appendix of the final version of the paper. **Justification for SFL & Different Definitions of Learnability**: Thanks for providing additional examples of learnability. This addresses the issues I raised. Given the above discussion, I believe most of the issues I raised were appropriately addressed by the authors. I will raise my score accordingly. Thanks again for the detailed answers and the additional experiments, and good luck with the paper

Authorsrebuttal2024-08-12

Response to ycoo

Dear Reviewer, Thank you for your timely response to our comments. We believe we have addressed the concerns you have raised in detail, including through additional experiments which confirm and strengthen our original findings. We also appreciate the increase in your score but were hoping for more significant support given your positive response.

Reviewer Ns9r7/10 · confidence 3/52024-07-09

Summary

This experimental paper proposes a UED method (SFL) for JaxNav, a continuous single- and multi-robot navigation task in a partially observable setting. The authors document the shortcomings of UED methods, (Domain Randomisation, Prioritised Level Replay, and ACCEL) on this partially observed, continuous action, continuous state setting. The authors provide a demonstration that the scoring mechanism used by the above UED methods is misguided. Their method, SFL, uses a “learnability score” which focuses learning on levels for which agents achieve success rates closer to 50%. To better focus on robustness and generalization, the authors develop an empirical CvaR measure of success for evaluation and comparison of methods.

Strengths

Scientists who are interested in developing and implementing the RL agents in real life will find this experimental paper important. The paper is technically sound and provides simulation support for the claims. The authors provide a nice discussion of the weaknesses of SFL (proposed version works only for binary outcome, deterministic environment). I have not gone over the github site; aside from the points made below, replicability is good. The learnability score is original (obvious after reading the paper, but perhaps not before). This score nicely takes advantage of the JaxNav environment which provides fast training of an RL agent. Related work is adequately cited and it is very clear how SFL differs from the 3 UED methods discussed.

Weaknesses

See questions below.

Questions

Lines 15-17. Suggest to delete spurious statements such as “We had tried our best to make current UED methods work for our setting before going back to the basics and developing our new scoring function, which took a few days and worked out-of-the-box. We hope our lessons and learnings will help others spend their time more wisely.” Don’t complain! The authors should make an effort to assist readers familiar with RL but not ACL methods by providing details in the appendix. For example in line 175, Algorithm 1 does not provide information on how \phi is updated. The update algorithm could be provided in the appendix. Figure 2 includes statements about “learnability” prior to a definition of learnability. Lines 154-157. Learnability has not yet been defined so statements about slight/no correlation with learnability are vacuous. It is not clear to reader whether learnability in these lines/Figure 2 is the same as the definition of learnability given later (and used by SLR). Lines 133-135 might better go in the later section on weaknesses/limitations. Line 214. The authors do not define “solvable.” The reader needs to know how you are operationalizing this term. Line 274-5. Unclear what “perfect regret” means.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Limitations are nicely discussed. It would be cool if the authors could comment on how SFL might be generalized to continuous outcomes and stochastic environments.

Reviewer gAKk7/10 · confidence 4/52024-07-12

Summary

This work proposes a new Unsupervised Environment Design (UED) method, called Sampling for Learnability (SFL), developed for navigation tasks. SFL upsamples environment configurations with high learnability scores, i.e., p*(1-p), where p is the success rate in a specific configuration. The paper highlights the failure of some of the state-of-the-art (SOTA) UED approaches on such tasks and argues that they cannot approximate learning agents' regret, unlike SFL. They evaluate SFL in JaxNav, a Jax-based single and multi-agent navigation benchmark they introduce, and MiniGrid. Their empirical results evidence that SFL is robust and can outperform Domain Randomisation (DS) and SoTA UED approaches in JaxNav and MiniGrid. The paper demonstrates these outcomes by following two evaluation protocols, respectively: 1) A risk-based protocol that they introduce, computing the conditional value-at-risk (CVaR) of the distribution of success rate based on randomly sampled levels, and 2) a common protocol in the UED literature that evaluates performance/success in hand-designed test sets, as they focus on complex, yet, solvable configurations.

Strengths

- This paper is well-written, clearly expresses the motivation and the gap in the literature, and illustratively analyses the failure of existing UED methods in navigation tasks. - Their contribution is a simple yet novel UED approach called SFL that explicitly addresses this failure mode. Their empirical results through two evaluation protocols highlight that SFL is robust (in terms of CVaR-based metrics) and outperforms existing approaches in expected return/success. - I agree with the authors' claim that existing UED methods claim robustness yet fail to quantify it accurately. Hence, the robustness evaluation protocol utilized in this work may be considered new in the UED literature.

Weaknesses

- Section 4.1 analyses MaxMC and PVL, popular UED score functions, in terms of their predictiveness of learnability. However, there is no demonstration of a similar analysis for SFL. Although SFL's scoring function is intuitive and it is not as challenging to guess what the scoring would look like, including such an illustrative comparison would support the claims in this paper. - The empirical results according to the proposed evaluation protocol for robustness (Figures 3a, 5a, and 7a) show that SFL outperforms the rest of the evaluated methods for alpha < 1 (100%). However, except in the multi-agent case (Fig 5a), the expected success rates indicate marginal or no improvement. In MiniGrid, the difference between DR and SFL drops quickly as alpha gets above 0.1 (10%). This is likely due to SFL not doing so well in easier levels compared to other baselines. Although this is expected, as the proposed metric results in upsampling learnable environments, not easy ones, I suggest a comparison of results in easy levels to conclude whether this is the case or not.

Questions

Table 4 showcases compute times for each evaluated approach. Most of these approaches have similar components, except maybe SFL's extra rollout phase. So why is SFL one of the fastest?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

- The major limitation of the proposed approach is the rollouts needed to compute the learnability score of randomly sampled levels. Section 4.2 indicates that 5000 levels are sampled, and n-step rollouts, where n is not specified, are generated to compute the scores. It would be more informative if the amount of computation and time spent in this phase were reported (in comparison to the usual curriculum learning task sampling phase). In addition, I wonder how SFL is one of the fastest in the single agent nav environment, as reported in Table 4 in Appendix F, despite having this additional phase.

Reviewer Su347/10 · confidence 5/52024-07-14

Summary

The paper introduces a new metric based on solve rates for evaluating the learning potential of tasks in a multi-task RL environment. It also presents a new evaluation protocol based on worst-case performance to better characterize the robustness of different methods. The work uses this protocol to compare their method with several Unsupervised Environment Design baselines and concludes that it outperforms the baseline. They also find that their UED baselines perform worse than Domain Randomization on this new evaluation metric, and conduct additional experiments to confirm and analyze these results.

Strengths

The paper includes many contributions, including a new method, a new evaluation protocol, and new benchmarks of existing methods. These contributions are open sourced and well documented in the appendix to support reproducibility. The work attempts to identify failure modes of a large class of regret-based UED algorithms which could help to move further ACL research toward more promising directions. These methods are effective and widely applicable, but their generalization outside of a small set of baselines has not been thoroughly studied. This work will become increasingly important as RL research moves on to more complex environments.

Weaknesses

The main weakness of the paper is the claim that domain randomization and their new heuristic outperform all UED algorithms, which is not entirely convincing from the results. The PLR hyperparameters used in this work are different from [1] and [2], for example the prioritization, temperature, and number of edits are all different. PLR in particular has several hyperparameters that need to be tuned for new domains, such as the buffer size, sampling temperature, and staleness coefficient. The authors do not explain whether they tuned the hyperparameters for the baselines or how they tuned the hyperparameters for their own method. I think even a small grid search over reasonable parameters that worked in other domains would help to make the comparisons throughout this paper more convincing. Another concern is that this work does not include any standard episodic return plots, which is the main metric of comparison in RL. Without these plots, it is difficult to tell whether their method truly outperforms the baselines, or whether their baselines are properly tuned. For instance, if the SFL algorithm performs better according to the CVaR evaluation but not on a standard test return plot, then it's debatable which method is more useful. That being said, [2] does compare the mean evaluation solve rate of UED baselines on Minigrid, and finds that Accel and PLR both significantly outperform DR. In this work, PLR severely underperforms DR using the same metric in Figure 3. This seems to be strong evidence that its hyperparameters or implementation are not correct. Even if we assume that the baselines are working as intended, I PLR and SFL have many differences which makes it unclear what change is actually leading to better performance. I think this work would benefit from additional experiments using the learnability metric in SFL as the prioritization metric in PLR, and possibly the alternative, using PVL and MaxMC as the selection metrics in the full SFL algorithm. Overall the writing and presentation was quite good, but there were a few notable issues. The paper somewhat confounds its description of PLR with Robust PLR in section 2.2.1. PLR trains on randomly sampled levels, while Robust PLR evaluates randomly sampled levels and only trains on levels from its replay buffer. Also, "learnability" is not clearly defined until 4.2, after it has been referenced many times. This is ok when discussing "intuitive notions of learnability" in a vague sense, but not when making statements such as "no correlation exists between learnability and regret score". [1] Jiang, Minqi, et al. "Replay-guided adversarial environment design." Advances in Neural Information Processing Systems 34 (2021): 1884-1897. [2] Parker-Holder, Jack, et al. "Evolving curricula with regret-based environment design." International Conference on Machine Learning. PMLR, 2022.

Questions

* Why did you not simply use your heuristic instead of PVL or MaxMC in prioritized level replay? * Why does there appear to be a discrepancy between the UED baseline solve rates reported in this paper and the ones reported in [2]? * On line 214 you describe sampling 10,000 solvable levels to evaluate on. Are these unseen levels (at least with high probability) or are they sampled from the training set? * You mention in the Limitations section that your method is restricted to Jax-based settings. Is the method limited to those settings, or just your particular Jax-based implementation?

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors adequately discuss the limitations in the paper and throughout the work. Their SFL method appears to be less general that the UED baselines that they compare against, but performs much better on the domains tested in this work. Simple methods on new benchmarks can serve as stepping stones to more general methods, so I do not believe this impacts their contribution.

Reviewer Ns9r2024-08-07

Thanks

Thanks for your response. I will hold to the rating of 7.

Reviewer ycoo2024-08-11

Thanks for the answer

Thanks for your detailed answer. I took some more time to think about the points you raised: **Different Definitions of Learnability**: I see that indeed rejecting impossible or already completely solved level is indeed the biggest factor in increasing performance. I would still argue however that I find the learnability definition not very principled - if what mentioned in the first sentence is true, many other scoring function which do the same in those extreme cases, and which chosoe a better heuristic in the remaining scenarios will outperform SFL. Isn't that the natural thing to try? **Hyperparameter choices / Ablations**: Can you clarify the sentence ' we have more insights now into which parameters are most important for SFL'. Which are these hyperameters, and more specifically, why is that the case? **More seeds and environments**: Thanks for adding more seeds and an additional environment. **Writing**: can you provide the most significant changes done to the writing? **Different values of the sampling ratio**: You say *increasing makes the agent train on more learnable levels, it may also decrease the diversity of the levels seen by the agent* - do you have any insight on the structure of the levels that are create by SFL, especially when $\rho$ is increased to 1? Many similar plots have been produced by other similar papers, such as Fig. 3 in the ACCEL paper (https://arxiv.org/pdf/2203.01302). **Average vs top-k learnability**: You should also compare this with the levels of learnability in the buffers of PLR and ACCEL for example, such as one can have a complete picture of where the advantage by using SFL is coming from. **What we did to make UED better** and **Different learnability definitions?**: Thanks for your answer In general, I agree with the authors when saying that SFL outperforms UED method in the environments considered. However, beside the remaining doubts above, I am still on the fence for this work. My main criticism comes from the fact that a new method (SFL) is proposed, but that there is no extensive justification on why this is happening. I believe it would be good to have more insight in the structure of the levels SFL prioritizes, and if there are other gaps SFL is actually filling beside excluding impossible or completely solved levels. Is learnability itself better, or any method which excludes the pathological type of levels I just mentioned would actually perform similarly? I believe this is an important point to address. Sorry for the delay in the answer, and I understand if you are not able to produce a response in time!

Authorsrebuttal2024-08-12

Response 1/2

Dear reviewer, thank you for your response, your continued engagement, and your commitment to improving our paper! Please find our responses below. # Hyperparameter choices / Ablations First, as illustrated in Figure 2 in the rebuttal PDF, SFL is robust to all of its hyperparameters, and suboptimal choices do not lead to catastrophic failure. That being said, the most important hyperparameters are: - $N$: How many levels to sample. Sampling more levels increases performance, as more levels gives SFL a larger sample to draw high-learnability levels from. - $K$: Buffer Size. In Jaxnav, we find a small buffer size is beneficial, and in Minigrid it is better to have a slightly larger one. This could relate to how easy the environment is (corresponding to how long it takes to learn a particular level). As JaxNav is harder than Minigrid (as it also involves low-level continuous control obstacle avoidance in addition to maze solving), it may be that training on each level more times is beneficial. By contrast, since Minigrid is easier, the number of episodes required in a level may be much less. - $T$: The buffer update period. If this value is too large, then performance degrades. This is because the learnability of the buffer, and therefore the usefulness of the levels, decreases as the agent trains on it. # Writing We kept the structure of our manuscript largely the same and made several local and low-level writing changes and fixes to ensure the writing flows better. One structural change we did make is to define learnability earlier on, in Section 4.1. We finally also added additional explanations and descriptions based on the reviews by yourself and the other reviewers. # Different values of the sampling ratio We agree such a plot would be useful, and will provide one in our revised manuscript. Since we are unable to share figures during the discussion period, we report the shortest path, number of walls, and solvability, averaged over training in the table below. We find that SFL has marginally longer shortest paths and marginally fewer walls. The SFL levels are also considerably more solvable. | Method | Shortest Path (Mean) | N Walls (Mean) | Solvable (Mean) | |:-----------|:-----------------------|:-----------------|:------------------| | PLR (PVL) | 4.74 (0.06) | 25.59 (0.19) | 0.89 (0.01) | | PLR (MaxMC) | 7.03 (0.49) | 22.63 (0.21) | 0.89 (0.02) | | SFL | 8.12 (0.02) | 21.98 (0.43) | 1.00 (0.00) | Aside from solvability, there is not a noticable difference in these metrics between PLR and SFL, despite SFL significantly outperforming PLR. Qualitatively, we find that in JaxNav, levels with high learnability tend to involve a lot of turning and intricate obstacle avoidance (as opposed to long paths). As such, the number of walls and shortest path length do not fully capture a level's difficulty. We will add illustrative examples of generated levels to our Appendix to demonstrate this point. # Average vs top-k learnability in UED: Thank you for suggesting this experiment, we have run the same analysis for ACCEL and PLR in JaxNav. We further compute the correlation between the score (e.g. MaxMC or PVL) and learnability for all levels in the buffer. We find that: - **The average learnability of levels in the PLR/ACCEL buffer is very low.** This is also true when selecting only the levels with the top 50 PVL/MaxMC scores. - There is no significant correlation between PVL/MaxMC and learnability. - Most of the levels selected by PVL and MaxMC can consistently be solved by the agent. The table below reports the learnability and success rates for levels within the PLR/ACCEL buffers averaged over training. At each evaluation step, we calculate the average and median values for the entire buffer and then average these values over training. Finally, we report the mean and standard deviation across three different seeds. | Method | Learnability (Mean) | Learnability (Median) | Success Rate (Mean) | Success Rate (Median) | |:-------------|:----------------------|:------------------------|:-----------------|:-------------------| | PLR (PVL) | 0.01 (0.00) | 0.00 (0.00) | 0.85 (0.01) | 0.96 (0.00) | | PLR (MaxMC) | 0.02 (0.00) | 0.00 (0.00) | 0.85 (0.03) | 0.95 (0.04) | | ACCEL (MaxMC) | 0.02 (0.00) | 0.00 (0.00) | 0.93 (0.01) | 0.97 (0.02) | | ACCEL (PVL) | 0.01 (0.00) | 0.00 (0.00) | 0.94 (0.02) | 0.97 (0.02) | This result further strengthens our findings (shown in Figure 2 of the original manuscript) that current UED score functions do not correspond to learnability. Instead, most levels in the UED buffers can already be solved 100% of the time. By contrast, as shown in our rebuttal, SFL consistently selects levels with high learnability.

Authorsrebuttal2024-08-12

Response 2/2

# Justification for SFL & Different Definitions of Learnability We believe the intuition for learnability is simple, easy to understand, and our score function works well empirically compared to prior approaches. However, we are not claiming that learnability is the be-all and end-all for score functions; we fully agree that there may be other variations of it that could perform better, and we are glad the reviewer raised this important point for future work. Furthermore, we would like to point out that the regret approximations used in current UED methods (e.g., MaxMC and PVL) are not particularly principled and are at best a loose approximation to regret. Indeed, in [1], six possible functions are trialled with their final choice based solely on empirical results. Prior work has also not investigated how closely these metrics correlate to true regret, and recent work has not investigated any alternatives to PVL and MaxMC. Further justification for our score function is presented below in three parts: **First, we wish to emphasise that our learnability definition is deeply connected to a large body of prior work in learning theory and curriculum development.** Intuitively, many curriculum methods (including SFL) aim to obtain experiences within the agent's "zone of proximal development", i.e., those that are neither too easy nor too hard [2]. Works such as [3] use hardcoded reward thresholds to define a goal that is of appropriate difficulty. Similarly, POET [4] also uses predefined reward thresholds to filter environments. SFL, meanwhile, directly uses the success rate of the agent on the level to compute the score. **Second, we believe our empirical results are a strong justification of SFL**, with the reviewer agreeing that "SFL outperforms UED". We would also note that most prior peer-reviewed work in UED uses two ([5, 6]) or three ([7]) environments, and we have strong empirical results in four. **Third, we have run several other variations on learnability, all of these assign a level a score of zero when $p=0$ or $p=1$** (in JaxNav): 1. Uniform sampling of levels with $0 < p < 1$. 2. Learnability linearly increasing from $p=1$ to $p=0$. 3. Learnability linearly increasing from $p=0$ to $p=1$. We find that variation 2. (which is an approximation of true regret) performs similarly to the default definition of learnability. Variation 1. performs worse than our approach, but still outperforms all UED methods (albeit by a smaller margin than SFL). Finally, variation 3. performs significantly worse, and we believe this is due to this score function prioritising easy levels, which provide less opportunity for the agent to improve. We will include these results in our updated Appendix. While we focused on $p(1-p)$, these results (and the previous results with different learnability peaks) show that our method performs robustly under other reasonable learnability estimates, demonstrating the potential of learnability in the UED setting and the robustness of SFL. While we agree that the exact choice of learnability metric is still an open question, we believe that all our claims are justified empirically, and our results make an impactful contribution to the UED community. **References** [1] Jiang, Minqi, et al. "Prioritized level replay." International Conference on Machine Learning. PMLR, 2021. [2] L. Vygotsky. Interaction between learning and development. Readings on the Development of Children, pages 34–40, 1978. [3] Carlos Florensa, David Held, Xinyang Geng, Pieter Abbeel. ICML, 2018 [4] Wang, Rui, et al. "Paired open-ended trailblazer (poet): Endlessly generating increasingly complex and diverse learning environments and their solutions." arXiv preprint arXiv:1901.01753 (2019). [5] Dennis, Michael, et al. "Emergent complexity and zero-shot transfer via unsupervised environment design." Advances in neural information processing systems 33 (2020): 13049-13061. [6] Jiang, Minqi, et al. "Replay-guided adversarial environment design." Advances in Neural Information Processing Systems 34 (2021): 1884-1897. [7] Parker-Holder, Jack, et al. "Evolving curricula with regret-based environment design." International Conference on Machine Learning. PMLR, 2022. # Conclusion Once again, we appreciate the reviewer's engagement in the review process, and their helpful suggestions. We believe these results further strengthen our paper.

Reviewer gAKk2024-08-12

Re: Rebuttal by Authors

Thank you for your effort in responding to my comments and questions and providing new results that support the validity of your work. My concerns have been thoroughly addressed, so I'll raise my score from 6 to 7.

Reviewer Su342024-08-13

I’d like to thank the authors for their very thorough response and comprehensive additional experiments, as well as apologize for not responding sooner. Thank you for performing a more thorough hyperparameter sweep, and for pointing out the difference in minigrid block budgets from prior work. After looking through JaxUED and Robust PLR again, I agree with the authors that UED is comparable to DR in the 50 and 60 block setting. That being said, I think the authors should also experiment in the more challenging 25 block setting where UED methods are most effective. Regardless, these new results provide convincing evidence that the proposed method is better than the SOTA in at least some settings. Than you also for running the requested ablations so quickly. The results seem inconsistent across hyperparameters and environments, but it does appear that SFL is generally better, and both the sampling and prioritization metrics positively impact performance. My main concern was that there seemed to be strong evidence that the evaluations in the paper were not fair. With these extensive new results and changes, I believe the contribution of this work is convincing, and I hope the authors will incorporate these new results in their paper as promised. I will raise my score to a 7.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC