Conditional Mutual Information for Disentangled Representations in Reinforcement Learning

Reinforcement Learning (RL) environments can produce training data with spurious correlations between features due to the amount of training data or its limited feature coverage. This can lead to RL agents encoding these misleading correlations in their latent representation, preventing the agent from generalising if the correlation changes within the environment or when deployed in the real world. Disentangled representations can improve robustness, but existing disentanglement techniques that minimise mutual information between features require independent features, thus they cannot disentangle correlated features. We propose an auxiliary task for RL algorithms that learns a disentangled representation of high-dimensional observations with correlated features by minimising the conditional mutual information between features in the representation. We demonstrate experimentally, using continuous control tasks, that our approach improves generalisation under correlation shifts, as well as improving the training performance of RL algorithms in the presence of correlated features.

Paper

Similar papers

Peer review

Reviewer n6Jo7/10 · confidence 3/52023-07-03

Summary

The paper mentions that current disentanglement techniques in RL require features to be independent and cannot disentangle correlated features. Therefore the paper proposes an auxiliary task for RL methods based on conditional mutual information (CMI) which can learn to disentangle observations with correlated features. The authors propose a conditional set which makes the features independent and details the process of learning conditionally independent representations. Since the conditioning set is based on the entire history of actions and past features, the authors claim that using the last one time-step of action and features is enough for good generalisation performance. The results section shows that the proposed method works well when reverse correlation is encountered during testing and also provides ablations to show the efficacy of different design choices.

Strengths

- The paper is very well written, and I found it very easy to follow through with detailed figures highlighting the conditioning set and the architecture. - The paper very clearly explains how to minimise the proposed CMI with explanations of how to sample from the joint and product of marginals distribution. - The proposed auxiliary task can be used on top of existing RL methods with just adding a discriminator and an adversarial loss. - The authors have shared their code for reproducibility. - There are clear ablation experiments to showcase the efficacy of different design choices made (correlation %, conditioning set, history length, CMID loss coefficient).

Weaknesses

- The experiments seem a bit limited as the proposed method is evaluated on four DMC tasks with only a couple of colour/size variations which raises concerns regarding the generalizability/scalability of this approach. The paper could have been more insightful with another domain with more complex correlations where longer history is required for the conditioning set. That would also have helped with understanding if the kNN permutations used would work well with increasing dimensions (N) of the feature space.

Questions

- When doing the forward pass to generate $\boldsymbol{z}_{t−1}$, can the authors explain why a momentum encoder is required instead of directly forward pass through encoder? - Line 188 should have $\boldsymbol{c}^n_{t’}$ instead of $\boldsymbol{c}^n_{t}$? - Can the authors showcase results on SAC without any data augmentation just to check how much difference does SVEA make on the DMC tasks?

Rating

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

Confidence

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

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes, the authors have pointed out limitations of their approach.

Reviewer oLSP7/10 · confidence 2/52023-07-06

Summary

The paper proposes an auxiliary task based on minimising mutual information in the learned latent representation. While several approaches have adopted similar loss functions for learning representations, the difference here is that the mutual information is conditioned on the history of past (latent) states and actions. This has the effect of disentangling correlated features, which improves performance when the agent interacts with a new task where those correlations no longer exist. Experiments on continuous control tasks illustrate that the method better generalises to tasks where these correlations are not present, whereas other approaches (also based on mutual information) fail to zero-shot transfer.

Strengths

The paper is well-written and easy to follow, while the experimental results clearly show the advantage of the method over other auxiliary tasks that have previously been proposed. Although I am not overly familiar with the literature in the space, I believe that the proposed method, while a variation of existing approaches, is novel. The auxiliary task is well motivated, and the implementation of it within a deep RL framework is elegant, which could likely make it easy for others to incorporate the idea into their own learning algorithms.

Weaknesses

Too much of the experiment section is in the appendix. As it stands, the main paper is not readable without reference to the appendix. I realise that space is an issue, but it would be best if the main text served as a standalone document. For example, there is no explanation of what reversed vs no correlation means (although I can guess). In the main text, only the pendulum task is described - the rest are not even briefly mentioned. To free up space, I recommend moving the discussion of history length and coefficient into the appendix instead. These can be briefly mentioned, but they serve as additional ablations and are not key to the story. Given Figure 13, the saliency maps in the main text seem a bit cherry-picked, since many of the other ones are not as clean. While the RL performance clearly shows a win, it would be nice to see more evidence that the auxiliary task promotes the kind of disentanglement that is expected. Would it not be possible to investigate this in the non-RL setting? For example, collect data and then simply apply the representation learning part alone to that data. The resulting representations could also then be evaluated using existing metrics from the representation learning literature (e.g. MIG score, D/C/I, FactorVAE, etc.) While the tasks are engineered to demonstrate the point of the paper (by synthetically causing entanglement), it would have been nice to see the approach tested in a task that wasn't quite as synthetic and hadn't been formulated precisely to showcase the strengths of the method. Minor: The intro contains many different examples; it would be clearer to stick with one running one consistently. The discussion in Section 4.1 may be inaccessible to someone in RL who is not familiar with graphical models, causal learning, etc. Given that space is an issue, it might be helpful to expand on some of these ideas in the appendix (e.g. explaining the concept of a collider). Perhaps Figure 1 could serve as a diagram where various concepts could be illustrated.

Questions

I'm a bit confused about how the method actually works. Clearly, from the empirical results, it works very well, but the thing that confuses me is that the conditioning set is based on the history of latent representations. But these latent representations change over time as the agent learns. So why does the kNN procedure actually work? If the representation changes over time, then won't the buffer contain "old" representations? For example, if the agent visited a state s early on, it would have some representation that's stored in the buffer. But later, if that same state is visited, the representation will be different (because the encoder has been updated since). I can see how things work if z is fixed, but I'm a bit perplexed that it works even though z is continuously changing. Any clarification here would help. Learning representations conditioned on history is reminiscent of recent work that seeks to learn a Markov representation (Allen et al., 2021). I'm wondering if there's a link there or if the ideas here are related to the Markov property in some way? Allen, Cameron, et al. "Learning Markov state abstractions for deep reinforcement learning." Advances in Neural Information Processing Systems 34 (2021): 8229-8241.

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

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The limitations of the method with regard to simple, synthetic correlations are discussed well. I wonder if one further limitation could be the reliance on the Euclidean distance metric. While it will likely be fine in the case of the state representation and continuous actions, I'm wondering specifically about the case of discrete actions, since the IDs attached to each action are arbitrary and so Euclidean distance between actions is meaningless.

Reviewer x11j8/10 · confidence 4/52023-07-06

Summary

This paper presents a novel approach to disentanglement in RL, focusing on the challenge of spurious correlations in high-dimensional observations. The authors argue that conventional disentanglement techniques, which rely on minimizing mutual information between features, can struggle with the task of disentangling correlated features, leading to a generalization failure when these correlations shift. To address this issue, they propose an auxiliary task termed Conditional Mutual Information for Disentanglement (CMID) which can be added to any RL algorithms. CMID constructs disentangled representations by minimizing the conditional mutual information between dimensions in the latent representation. The authors experiment with continuous control tasks and show that CMID enhances training and generalization performance of base RL algorithm SVEA, and outperforms other state-of-the-art baselines, supporting its efficacy in handling correlated features and correlation shifts.

Strengths

This paper exhibits several significant strengths that enhance its value: * The paper is meticulously written with a clear and organized structure that allows for easy comprehension of its contents. Each claim is sufficiently justified and is coherently interconnected to form a compelling narrative. * The authors present a versatile and elegant method, CMID, that can be conveniently integrated with any model-free RL algorithm. This broad applicability enhances the practical relevance of the proposed method. * The presented results offer persuasive evidence for the effectiveness of the proposed method. Notably, improvements in both training performance and zero-shot generalization are demonstrated, indicating successful disentanglement of information. * The analysis of the agent's attributions provides a valuable qualitative insight into the learned representations. This analysis adds depth to the study and helps to illustrate the inner workings and benefits of the proposed approach.

Weaknesses

* One of the main drawbacks of the proposed method is the computational overhead, particularly due to the loop over the features during the computation of the adversarial loss. While this is a natural consequence of the chosen method, it would be beneficial for the authors to quantify this overhead. * The paper could also benefit from a more direct qualitative or quantitative evaluation of the disentanglement effectiveness. The provided experiments successfully showcase improvements in terms of reward and zero-shot adaptation, but they do not directly assess how well the disentanglement itself is working. For instance, measuring the impact of changing specific factors (like the color or length of a cartpole) on the N feature would provide a clearer picture of the disentanglement capability of the proposed method and reinforce the authors' claims. This additional experiment, while not necessary for rebuttal, could greatly strengthen the paper's contribution and its implications in practical settings.

Questions

* The paper mentions the use of a momentum encoder to compute the conditional set. Could the authors elaborate on why this is necessary? Have the authors considered or experimented with using only one encoder and stopping gradients? * The authors note that in theory, the conditioning set should include the full history, despite the markov property. This concept could use further elaboration, as it's not entirely intuitive why this would be the case. Could the authors provide further explanation on why the full history is still necessary in the conditioning set?

Rating

8: Strong Accept: Technically strong paper, with novel ideas, excellent impact on at least one area, or high-to-excellent impact on multiple areas, with excellent evaluation, resources, and reproducibility, and no unaddressed ethical considerations.

Confidence

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

Soundness

4 excellent

Presentation

4 excellent

Contribution

3 good

Limitations

The authors adequately addressed current limitations of this work.

Reviewer 42Lr5/10 · confidence 3/52023-07-07

Summary

Reinforcement learning often suffers from feature correlations from the limited number of data and feature coverage which hinders the generalization to other environments. To this end, this paper proposes to add an auxiliary task for RL algorithms to learn a disentangled representation of high-dimensional observations with the correlated features by minimizing conditional mutual information between features in the representations. The conditional mutual information is minimized by the density ratio trick using batch permutation and discriminator. This method improves robustness of RL algorithms.

Strengths

* The motivation of this paper is very clear and the idea is easy to follow. * The experiment setting is quite reasonable.

Weaknesses

* The experiment results only focus on SVEA. Authors should add experiment results with other RL algorithms. * The related works for RL algorithms with disentangled representations are not enough.

Questions

Check weaknesses

Rating

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

Confidence

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

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

Yes

Reviewer x11j2023-08-11

2 / 2 I appreciate the authors' response and the clarification they provided in addressing my questions. Taking their explanations into account, I continue to find this paper to be valuable.

Reviewer oLSP2023-08-12

Thanks to the authors for their response, which clears up some of my confusion. I'm happy to stand by my initial evaluation and recommend the paper for acceptance to the conference

Reviewer n6Jo2023-08-13

I thank the authors to provide clear answers to the questions and providing additional ablation/results. Overall, I am happy to increase my score and recommend this paper for acceptance.

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC