Seeing is not Believing: Robust Reinforcement Learning against Spurious Correlation

Robustness has been extensively studied in reinforcement learning (RL) to handle various forms of uncertainty such as random perturbations, rare events, and malicious attacks. In this work, we consider one critical type of robustness against spurious correlation, where different portions of the state do not have correlations induced by unobserved confounders. These spurious correlations are ubiquitous in real-world tasks, for instance, a self-driving car usually observes heavy traffic in the daytime and light traffic at night due to unobservable human activity. A model that learns such useless or even harmful correlation could catastrophically fail when the confounder in the test case deviates from the training one. Although motivated, enabling robustness against spurious correlation poses significant challenges since the uncertainty set, shaped by the unobserved confounder and causal structure, is difficult to characterize and identify. Existing robust algorithms that assume simple and unstructured uncertainty sets are therefore inadequate to address this challenge. To solve this issue, we propose Robust State-Confounded Markov Decision Processes (RSC-MDPs) and theoretically demonstrate its superiority in avoiding learning spurious correlations compared with other robust RL counterparts. We also design an empirical algorithm to learn the robust optimal policy for RSC-MDPs, which outperforms all baselines in eight realistic self-driving and manipulation tasks.

Paper

References (100)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer 27Kb6/10 · confidence 3/52023-07-07

Summary

This work aims to make RL robust to distribution shifts by limiting their reliance on spurious correlations between state features. This is achieved through a new RL algorithm designed in a variant of Robust MDPs extended to include a more structured uncertainty representation.

Strengths

* This paper addresses a serious underlying flaw in the standard approaches for Robust MDPs, in that the uncertainty set is usually artificial an no alined to the ways we need our policies to be robust.

Weaknesses

The language is often confusing, for instance the term "spurious correlations" is often used to refer specifically to spurious correlations between state variables. This is very confusing as it overloads the term, and makes it hard to refer to other spurious correlations (such as a spurious correlation between the agents actions and the reward or another agents actions). Better to say "Spurious state correlations". On a similar note "semantic uncertainty" is a confusing term that wasn't explained until deep into the paper. It sounds like uncertainty **about** semantics (which is how it has been used in the literature), but it is meant to be uncertainty that is not just a norm-ball perturbation over the transition function. A better term would be "structured" uncertainty. * Other more minor confusing wordings make the paper more difficult to read than it needs to be: * I don't know what it means for portions of the state to "not have causality" * I don't know what it means for an uncertainty set to be "shaped by the unobserved confounder and sequential structure of RL" * I don't know what it means for some approach to be "superior in breaking spurious correlations" * There is a typo in : "Despite various types of uncertainty have been investigated in RL" In addition it's unclear how sensitive the approach is to the degree of robustness. It's a famous issue of RMDPs that the degree of robustness has to be carefully selected, too large and you have an overly conservative policy, and too small you get no robustness. Giving that part of the argument for this approach is alleviating that problem (as shown in Figure 3), a sensitivity analysis here would greatly improve the paper, as a main draw away from RMDPs is that they are very sensitive to the robustness parameter.

Questions

How sensitive is the approach to changes in B% and K?

Rating

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

Confidence

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

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

It is important to discuss somewhere how hyper-parameters (such as k and B%) were chosen, and could be chosen in a new domain.

Authorsrebuttal2023-08-14

Thanks for your insightful suggestions!

Dear reviewer, Thank you once again for investing your valuable time in providing feedback on our paper. Your insightful suggestions have led to significant improvements in our work, and we look forward to possibly receiving more feedback from you. Since the discussion period between the author and reviewer is rapidly approaching its end, we kindly request you to review our responses to ensure that we have addressed all of your concerns. Also, we remain eager to engage in further discussion about any additional questions you may have. Best, Authors

Reviewer EdL36/10 · confidence 3/52023-07-08

Summary

The paper proposes a state-confounded (SC-) and a robust-state confounded (RSC-) MDP formulation to account for setups where a confounder satisfying the backdoor criterion confounds the states. The RSC-MDP setup assumes that the confounder lies in an uncertainty set which is part of the MDP parameters, and the aim is to learn a value function that is the lower bound of the different value functions arising out of different instances of the RSC-MDP at different values of the confounder within its uncertainty set. They propose an empirical algorithm to approximate the effect of perturbing the confounder by perturbing single dimensions of the states and setting them to values present in ‘otherwise nearby’ states (defined by distance between the two states in dimensions other than the dimension being edited). The paper claims that this indirectly estimates the effects of perturbing the confounder. They learn a graph-based model that is supposed to be a causal model, to generate new transitions (given a state and action s_t, a_t, generate s_t+1, r_t) based on the perturbed states and (original) action pairs. They augment SAC by randomly replacing part of the sampled batches with the generated transitions and learn as usual, and show that this recipe is robust to the spurious correlation in the nominal envs they define. All experiments are in the state space (non-vision inputs).

Strengths

1. The paper studies an important problem and proposes a set of benchmarks with handcrafted spurious correlations in a CARLA environment and in a robosuite environment which could be useful for more future works to study spurious correlations in RL. 2. The RSC-MDP formulation could also be useful, although I’m not sure how much the notation and proofs scale to an arbitrary number of confounding variables.

Weaknesses

1. The paper structure is somewhat confusing to me: The MDP formulations and the empirical algorithm don’t seem to have much of a connection and seems like both were developed independently - please correct me if I’m missing something here. It would be great to further clarify exactly how the algorithm is helping to solve the robust SC-MDP. 2. The notion of semantic uncertainty is not very precise or clear (I don’t think there’s any references either), and could be removed in my opinion since the RSC-MDP formulation can simply use the “uncertainty set” phrasing. 3. The appendix suggests that the training environments have perfect spurious correlations: the correlation isn’t broken in even a few instances of the training environments, so for eg. there would never be a (non-generated) transition in the SAC buffer that would have the brightness value low in its state space value, in the case that the training envs are following the correlation regime where brightness and traffic are both high during the day. It’s unclear to me then, why the causal model would ever generate a state value that doesn’t also have the brightness value set to high (same logic for the perturbation procedure). Am I missing something else here? Or are the training envs set up such that an agent will train on two kinds of envs at once - night time with less traffic and day time with more traffic, and so it s possible to see samples with the brightness value set to both low and high across otherwise nearby states? It would be great to clarify what exactly is the train and test env distribution and what is expected of the graphical causal model - is it to generate counterfactual transitions (counterfactual to what?) and conclusively show that it is actually doing that by visualizing or plotting some property of the samples it is generating. It's possible that I have misinterpreted something here, and I'm willing to raise my score if the authors' response clarifies some of these questions. 4. I would also suggest mentioning relevant related work explicitly aimed at resolving causal confusion in online [1] and offline RL [2] (and comparing to the closest version if applicable). 5. It would be great to include the training curves for the results in Table 1 as well, since that gives insight into whether other methods simply converge later to a similar highest reward, or whether they are entirely limited in their ability to converge to a similarly high reward as the best performing method. [1] Resolving Causal Confusion in Reinforcement Learning via Robust Exploration . Clare Lyle, Amy Zhang, Minqi Jiang, Joelle Pineau, Yarin Gal. ICLR Self-Supervised RL Workshop 2021. [2] Can Active Sampling Reduce Causal Confusion in Offline Reinforcement Learning? Gunshi Gupta, Tim G. J. Rudner, Rowan Thomas McAllister, Adrien Gaidon, Yarin Gal, NeurIPS Offline RL Workshop 2022, CleaR 2023

Questions

1. How is the experiment to answer R3 designed? It's unclear how to test the w/o P^{c} case since the confounder isn't known anyway? 2. As stated previously, it would help to get more insights into how novel the generated transitions from the causal model really are. 3. Have you observed or quantified the robustness-performance tradeoff mentioned in line 263, which will lead to a performance drop in some envs? I expect the perturbation procedure of states to only work for specific kinds of envs, so it would be good to show the failure cases as well.

Rating

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

Confidence

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

Soundness

2 fair

Presentation

2 fair

Contribution

2 fair

Limitations

Given how heuristic the technique of generating perturbed states is (swapping single dimensions between otherwise nearby states), the paper should properly discuss the many challenges of scaling this approach to a high-dimensional state space like that in visual domains.

Authorsrebuttal2023-08-14

Thanks for your insightful suggestions!

Dear reviewer, Thank you once again for investing your valuable time in providing feedback on our paper. Your insightful suggestions have led to significant improvements in our work, and we look forward to possibly receiving more feedback from you. Since the discussion period between the author and reviewer is rapidly approaching its end, we kindly request you to review our responses to ensure that we have addressed all of your concerns. Also, we remain eager to engage in further discussion about any additional questions you may have. Best, Authors

Reviewer EdL32023-08-16

Thanks for your response and new results

Dear Authors, Thankyou for adding the dynamics visualization as well as another baseline comparing to the proposed method. I think the comparison helps to highlight how the proposed method tackles a problem that is not addressed by the baseline - that of compositional generalization (please correct me if I'm conflating different things). I had the following question before I can update my score: Is the following correct? To generate a training example for the causal graph model the following procedure will be followed: We will take a transition (presumably from a daylight driving scenario) with day-light=high, and traffic=heavy, (and the other dimensions being set to some values), and assuming that we decide to perturb the day-light dimension of the current state, we look for states in the buffer where the day-light value is very different, however the other values are similar (therefore likely traffic=heavy, and day-light=low, and other dimensions being similarly valued). Now the s_{t+1} that is used to supervise the causal model output is still the same as that in the original transition (presumably day-light=high and traffic=heavy). I'm confused why the causal model should be expected to predict that a day-light heavy state transitions to a state that is day-light low -- unless the causal graph completely ignored the day-light dimension entirely. Getting the model to ignore the day-light dimension entirely is what I assume is being enforced by the sparsity loss - which leads me to the hypothesis that essentially what the causal graph is doing is helping to enforce the principle of "looking or depending on as few input dimensions as possible". This is important to note, as there is prior work [1] testing this idea for imitation learning by using some sort of dropping out of the input representation of a policy, (they don't make the assumption of operating in the state space). This also suggests that a similar trick might work if for example you removed the causal graph and perturbation mechanism entirely and trained a SAC policy with a input dropout mechanism on the state space inputs since it might enforce the same invariance to irrelevant dimensions. I think comparing the proposed method to this simple baseline will greatly help to refute this claim, as well as further explain what benefit the causal model is really bringing. I do think the proposed method has potential merit, but given the complexity and added assumptions of the additional components in the proposed method, I would like to be sure by dissecting rigorously where the gains are coming from. Thanks for your responses so far, and looking forward to engaging further. [1] Object-Aware Regularization for Addressing Causal Confusion in Imitation Learning Jongjin Park, Younggyo Seo, Chang Liu, Li Zhao, Tao Qin, Jinwoo Shin, Tie-Yan Liu

Authorsrebuttal2023-08-17

Response to follow-up questions

Thank you for engaging in the discussion and providing insightful feedback! We provide new experiments as well as analyses to answer your questions. ### **Q1: The proposed algorithm (RSC-SAC) tackles compositional generalization** Yes, we totally agree that there is a strong connection between compositional generalization and breaking spurious correlation. Actually, In section 5.1, we propose two kinds of environment settings of spurious correlation: 1. **Distraction correlation**: Between task-relevant and task-irrelevant features. The task-irrelevant feature is a distractor and should be ignored by the policy (random dropping may solve this); 2. **Composition correlation**: Between two task-relevant features. This exactly describes the compositional generalization setting, where the testing environments contain new combinations of task-relevant features (random dropping cannot help). ### **Q2: Correctness of the example of daylight driving scenario** The reviewer accurately addresses most aspects of the example, with the exception of the last sentence. After having the perturbed $\{s_t, a_t\}$ (e.g., traffic=heavy, day-light=low), we infer $s_{t+1}$ from our causal model with counterfactual generation. We expect a new $s_{t+1}$ by imagining a different value in $s_t$. In experiments, we observe that most $s_{t+1}$ are different from the original one and have traffic=heavy and day-light=low. ### **Q3: Adding experiments -- prior work OREO [1] as a baseline** We evaluate the performance of OREO [1] with different ratios of dropping ($\alpha$) on the Brightness (Distraction) and Behavior (Composition) environments. The results are shown below: |Method|Ours|OREO ($\alpha$=0.1)|OREO ($\alpha$=0.2)|OREO ($\alpha$=0.3)|OREO ($\alpha$=0.4)|OREO ($\alpha$=0.5)| |:-:|:-:|:-:|:-:|:-:|:-:|:-:| |Brightness (nominal)|0.92±0.31|**0.973 ± 0.239**| 0.739 ± 0.327| 0.371 ± 0.229| 0.26 ± 0.243| 0.207 ± 0.197| |Brightness (shifted)|**0.99±0.11**|0.891 ± 0.153| 0.562 ± 0.18| 0.256 ± 0.087| 0.182 ± 0.087| 0.128 ± 0.059| |Behavior (nominal)|**1.06±0.07**| 1.04 ± 0.104| 0.989 ± 0.122| 0.855 ± 0.259| 0.715 ± 0.282| 0.482 ± 0.224| |Behavior (shifted)|**1.02±0.09**| 0.517 ± 0.208| 0.541 ± 0.121| 0.553 ± 0.107| 0.509 ± 0.169| 0.366 ± 0.137| 1. Our method outperforms OREO in both Brightness (shifted) and Behavior (shifted) environments. We find that OREO indeed achieves robustness (but still worse than ours) in Brightness, which only contains the distraction correlation. 2. The advantage of our method (swap values of some dimensions of states) over OREO (drop some dimensions of states) could be explained from two aspects: * **Our method has compositional generalization, while OREO does not.** Swapping dimensions within the state creates new compositions of features that can address both compositional generalization and distraction issues. However, dropping information (OREO) can only deal with the distraction issue but does not make policy generalizable to unseen feature combinations. * **Dropping dimension of states loses information.** [1] focuses on image input which has a spatial structure -- still contains enough information after dropping some dimensions. However, in our state input setting (each dimension has semantic meaning), dropping dimensions of the state could cause severe information loss. The evidence is that increasing the ratio of dropping dramatically degrades performance. [1] Object-Aware Regularization for Addressing Causal Confusion in Imitation Learning Jongjin Park, Younggyo Seo, Chang Liu, Li Zhao, Tao Qin, Jinwoo Shin, Tie-Yan Liu

Reviewer EdL32023-08-17

Thanks for your response and explanation of the contribution of the causal graph

Dear Authors, Thank you for your response answering my questions and promptly adding the dropout baseline to help with the analyses. To confirm that I understood the response correctly: 1. In the response to Q2 you say that a new s_t+1 would be expected when the s_t is perturbed, does that mean the causal graph is never trained on the perturbed samples ie. it is trained on the non-perturbed transitions in the dataset, and then used for inference on the perturbed samples? (since we wouldn't have the correct supervision for s_t+1 for them) What is the intuition for why the causal model would be correct on inference on the perturbed samples with novel combinations of the features (If not for the sparsity inducing bias)? My follow-up question would be, how much is the performance reliant on the accuracy of the different components of the transition (reward and next state, or even just the combination of the perturbed s_t and original a_t existing in the buffer) - are both reward and next state prediction accurate? If the tasks are sparse reward I suppose most of the transitions just have a zero reward being predicted? 2. About the OREO baseline : Thanks for adding this. Is this the exact method from the OREO paper with the VQVAE component (which IMO is not needed since we are not operating in the image space here and already have all the relevant features as a compact input space), or do you mean that you used their dropping out mechanism on your input space and call that baseline OREO in the paper? In the meanwhile I have raised my score to Weak Accept in light of the new baselines and the added clarifications. Thanks!

Authorsrebuttal2023-08-17

Response to follow-up questions

Thanks for increasing the score! In response to the clarification questions: 1. The causal graph is not trained on perturbed samples due to the absence of the ground truth. The intuition behind the correct prediction is indeed the **sparsity inducing bias**, as pinpointed by the reviewer. Our hypothesis posits that the subsequent state holds greater significance than the reward. This is based on the observation. We hypothesize that the next state is more important than the reward since in most cases the reward is not influenced by the changing of one dimension of the state. 2. In our implementation, we just use the dropping-out mechanism in this baseline without using VQ-VAE. To be more accurate, we will name this method 'state dropout' instead of 'OREO'.

Reviewer EdL32023-08-17

Follow-up question

I see, in that case I have two questions/points of observation: 1. I'm confused about why a sparsity-inducing loss helps compositional generalization. To me it sounds like all tasks are essentially still challenging the distractor-avoiding skill of an agent, and the compositional case could probably be reformulated as a distracting case. Could you provide an example of what the spurious correlation is between the inputs (state dimensions) and outputs (action dimensions of a policy) in the compositional data, without the notion of distractors coming in? 2. I don't think random dropout would be the closest analogy here: apologies if my previous comparison to OREO added confusion into the mix - I wanted to refer to their insight that dropping out meaningful parts of the inputs or representations could help with robustness of a policy. In this case, I think a sparsity loss is what would help to understand if a learned/guided version dropout by itself is largely responsible for the performance improvements. This could be implemented either as a simple matrix multiplication before the first layer of a policy model with trainable matrix weights penalized by L1 loss (similar to what is happening in the causal graph), or even an L1 loss on the first or second layer MLP weights of the policy, to force it to attend to as few input dimensions as possible (again similar to what the causal graph is doing). Lastly, thank you for engaging in the discussion and helping to better understand the workings of your method.

Authorsrebuttal2023-08-18

1. **The compositional generalization mainly comes from two designs**: (1) (the major one) We swap the dimensions within the state, which creates new compositions of useful features that are unseen in the original training data. (2) We infer a reasonable $s_{t+1}$ according to the modified $s_t$, in which the sparse causal model could provide an accurate prediction. 2. **One example of the compositional setting** is the Door environment (Figure 8 in Appendix), where there is a spurious correlation between **the distance between the robot and the door** and **the position of the handle**. Both of these features are important for the robot to open the door. Specifically, the training env has [DoorDist=far, HandlePos=high] and [DoorDist=near, HandlePos=low]. After training policy on such a dataset, we observe that the robot always tries to reach the [high] position when the [DoorDist=far] even if it observes [HandlePos=low]. This can be explained as a failure of compositional generalization due to the spurious correlation between state and action. 3. We apologize for the misunderstanding of the baseline setting mentioned by the reviewer. So we add a new baseline (denoted as Sparse-SAC) that has an **additional L1 loss $\alpha \|W\_{1}\|\_1$** during the optimization, where $W\_{1}$ is the parameters of the first MLP layer of the policy and value networks and $\alpha$ is the weight. The results are shown below. Type of spurious correlation|Environment|Ours|SAC (non-robust)|Sparse-SAC ($\alpha$=1)|Sparse-SAC ($\alpha$=10)|Sparse-SAC ($\alpha$=100)|Sparse-SAC ($\alpha$=1000)| |:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:| |Distration|Brightness(nominal)|0.92±0.31|1.00±0.09|1.01±0.10|1.02±0.10|**1.03±0.10**|0.90±0.29| |Distration|Brightness(shifted)|**0.99±0.11**|0.56±0.13|0.69±0.2|0.87±0.18|0.83±0.13|0.72±0.29| |Composition|Behavior(nominal)|**1.06±0.07**|1.00±0.08|0.98±0.115|1.01±0.08|0.95±0.13|0.67±0.21| |Composition|Behavior(shifted)|**1.02±0.09**|0.13±0.03|0.16±0.032|0.15±0.04|0.10±0.02|0.08±0.02| Two important findings based on the results: * The sparsity regularization can improve the robustness of SAC in the setting of distraction spurious correlation (Brightness). This verifies the hypothesis of the reviewer that sparsity can address the distraction issue. * The sparsity regularization does not help with the composition spurious correlation (Behavior), which indicates that purely using sparsity regularization can not improve compositional generalization. We will add all these findings in the paper to illustrate the advantage of our method. We really appreciate the reviewer for proposing this interesting and effective baseline to improve the quality and presentation of our work, making the contribution more convincing and clearer.

Reviewer EdL32023-08-21

Thanks for adding sparsity loss experiments

Dear Authors, Thanks again for your engagement, and I'm glad that the paper has benefitted from it by including baselines that further help to dissect and highlight the key advantages of the proposed method. I still have some reservations/unanswered questions about why the causal graph model can generalize to give the correct prediction for a perturbed s_t. You also seem to attribute the accuracy of it's predictions to the sparse-ness of the causal graph - which to me suggests that this should then be achievable through the sparsity loss as well, unless this kind of sparsity is adaptive in some way and doesn't always choose to ignore the same input dimension always (thus making it different from the distractor correlation regime). I am inclined to increase my score to reflect a stronger recommendation towards acceptance if the authors can provide an answer to this (or add analysis to demonstrate or refute the above point - however I understand there is a time constraint and more experiments are strictly optional). Thanks!

Authorsrebuttal2023-08-21

Thanks for your question "Why the causal graph model can generalize to give the correct prediction for a perturbed s_t?" Unfortunately, we currently **do not** have theoretical proof of why the causal graph can accurately predict $s_{t+1}$, so we would like to empirically illustrate some cases where the prediction of our causal model is more accurate than using a neural network without causal. In general, the advantages are induced by the Independent Causal Mechanism (ICM) [1] that helps address the out-of-distribution issue. Assuming there are spurious correlations between two dimensions $s_t^1$ and $s_t^2$ of state $s_t$, we swap the values of two samples for the dimension $s_t^1$. * If we use a dense neural network (without causal graph) to predict $s_{t+1}$, we may get unpredictable results since the perturbed $s_t$ (a new composition of $s_t^1$ and $s_t^2$) is an out-of-distribution input to the neural network. In our experiments with the dense neural network, the predicted $s_{t+1}$ usually contains invalid values or the same value as the original samples. (We cannot provide the figure since we are not allowed to update the pdf file.) * If we use a sparse causal graph to predict $s_{t+1}$, we can benefit from the Independent Causal Mechanism (ICM), which learns independent sub-causal graphs between $s_t$ and $s_{t+1}$. Taking one ideal case with ICM as an example, when dimensions 1 and 2 are independent of each other in the causal graph (i.e., the causal graph is $s_t^1 \rightarrow s_{t+1}^1, s_t^2 \rightarrow s_{t+1}^2$), the prediction of $s_{t+1}^1$ is independent of $s_t^2$ and the prediction of $s_{t+1}^2$ is independent of $s_t^1$. Benefited from the decoupling, our causal model can accurately predict $s_{t+1}^1$ (only based on $s_t^1$) and $s_{t+1}^2$ (only based on $s_t^2$) since the dataset contains useful $s_t^1$ and $s_t^2$ separately. Even in non-ideal cases where dependency exists, we can still (empirically) benefit from the sparse connection to avoid $s_{t}$ becoming an out-of-distribution sample to the prediction model to increase the accuracy. [1] Parascandolo, Giambattista, Niki Kilbertus, Mateo Rojas-Carulla, and Bernhard Schölkopf. "Learning independent causal mechanisms." In International Conference on Machine Learning, pp. 4036-4044. PMLR, 2018.

Reviewer uQEX6/10 · confidence 3/52023-07-09

Summary

This paper aims to address the spurious correlation challenge that arises in RL. Such correlation is typically useless to decision-making but may be learned by agents, which leads to failure in applying to unknown test cases. To this end, the authors proposed a novel RSC-MDP framework that models the spurious correlation challenge with an unobserved confounder. The authors then justify that previous robust MDP methods may fail on the RSC-MDP framework due to inaccurate uncertainty modeling. The authors proposed a heuristic method to handle the RSC-MDP challenge authors then conducted extensive experiments to compare the proposed method and SOTA baselines for robust RL. Experiments justified that the proposed method has better performance for the proposed spurious correlation attacks.

Strengths

1. The paper is very well written, with a clear explanation of motivation and well organized problem setup and experimental design. 2. The authors provide theoretical justification that the proposed robustness is milder and more accurate than the traditional robust MDP. 3. The authors provide novel experimental design to assess the effect of spurious correlation in the state space of RL. The proposed baseline has the potential to be adopted by future study.

Weaknesses

1. In the proposed driving example, the confounder issue spans over multiple steps of interactions, whereas the latent variable c introduced in RSC-MDP only affects one step, and each step is affected independently (i.e., $c_t$ is independently generated for all $t$). In reality, the brightness and traffic density do not normally get perturbed per each step of the transition. Do the proposed RSC-MDPs cover the targeted spurious correlation problem? 2. Seems that the confounding within stat is not affecting the distribution of transition; that said, the observation still follows the distribution of trajectories after marginalizing the confounder $c$. To me, the spurious correlation challenge of RSC-MDPs is more of a robustness challenge than a confounding challenge to me, where the offline trajectory distribution is typically different from the interventional transition dynamics. How does the spurious correlation within each state affect the learning of transition P(s_{t+1} | s_t, a_t), hence affecting regular RL algorithms? 3. Could authors provide more reasoning on the proposed heuristic perturbation in approximating the true underlying perturbation of latent variables? Intuitively it seems to align well with the motivative driving example, but it seems that the perturbation proposed in equation (6) introduces extra correlations within states. In particular, sample $s^i_k$ selected to replace $s^i_t$ is based on a rule that explicitly involves $s^j_t$ for $j\neq i$. Do we worry that such introduced correlation ruins training for similar reasons of the targeted spurious correlation? Misc: line 81: discounted infinite-horizon -> finite horizon? line 220: $\sigma_2 \in (3/4, 1]$ according to the proof.

Questions

See Weaknesses.

Rating

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

Confidence

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

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

N.A.

Authorsrebuttal2023-08-14

Thanks for your insightful suggestions!

Dear reviewer, Thank you once again for investing your valuable time in providing feedback on our paper. Your insightful suggestions have led to significant improvements in our work, and we look forward to possibly receiving more feedback from you. Since the discussion period between the author and reviewer is rapidly approaching its end, we kindly request you to review our responses to ensure that we have addressed all of your concerns. Also, we remain eager to engage in further discussion about any additional questions you may have. Best, Authors

Reviewer 3Qte6/10 · confidence 4/52023-07-10

Summary

The paper studies how to develop more robust reinforcement learning algorithms when spurious correlation exists in the observation space. The paper studies the problem from a causal perspective and present robust state confounded MDP as the problem formulation. The paper proposes an algorithm to solve this problem formulation by learning the structural causal model. The paper demonstrate experimentally that the proposed method is more robust to shift in the environment compared to baselines.

Strengths

The paper is mostly extremely well written and easy to follow. The structure of the paper makes sense, which first discusses the formulation of the problem in the form of robust state confounded MDP and comparison to existing formulation. The algorithm the paper proposes is intuitive and the improvements over the baseline in the experiment section are clear.

Weaknesses

The discussion around related work focuses on a few most relevant papers and does not address the broader literature if I am not mistaken. The problem of spurious correlation is known within the community and various prior works [1, 2]. Please see [2] for example, which also learns the dynamics model to alleviate the issue of spurious correlation. These prior works tackle this problem even though they did not propose the same robust state confounded MDP that the paper proposes. I would have liked to see a more thorough comparison with the broader literature in a related work section. On page 3, the paper writes “people usually prescribe the uncertainty set of the transition kernel using a heuristic and simple function rho with a relatively small sigma”. However, because of the lack of a related work session, I do not see evidence for this statement. [1] https://ben-eysenbach.github.io/rpc/ [2] https://arxiv.org/abs/1909.11373

Questions

Please see my questions about related work

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

2 fair

Contribution

3 good

Limitations

Please see my questions about related work. I would have like to see the limitations that the method currently only work for state-based policies more prominently displayed and discussed.

Reviewer 3Qte2023-08-13

Thank you for updating related works

I am keeping my score as is, since I am not entirely convinced the related work section is well written just yet. For example, on page 16, there is this sentence "The proposed RSC-MDPs can be regarded as addressing the state uncertainty since the shift of the unobserved confounder leads to state perturbation. In contrast, RSC-MDPs consider the out-of-distribution of the real state that will directly influence the subsequent transition in the environment, but not the observation in POMDPs and SA-MDPs that will not directly influence the environment." I am quite confused what the "In contrast" refers to, since the previous sentence also discusses RSD-MDPs.

Authorsrebuttal2023-08-13

Thanks for engaging in discussion

Thank you for engaging in discussion with us and pointing out this question. To directly answer the reviewer's question, 'in contrast' refers to other prior works that also address state uncertainty as our RSC-MDPs. As the reviewer suggested, we have revised the related work section. We hope the following sentences make sense to the reviewer. > The proposed RSC-MDPs can be regarded as addressing the state uncertainty since the shift of the unobserved confounder leads to state perturbation. **In contrast to prior works which also address state uncertainty**, RSC-MDPs consider **distribution shift** of the real state that will directly influence the subsequent transitions in the environment, ~~but not~~ instead of the observation in POMDPs and SA-MDPs that will not directly influence the environment **but implicitly influences the policy**."

Reviewer 27Kb2023-08-18

Response to Rebuttal

Thank you for the clarifications. Your comments both here and to the other reviews largely address my concerns, and I will be increasing my score accordingly.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC