Distributional Model Equivalence for Risk-Sensitive Reinforcement Learning

We consider the problem of learning models for risk-sensitive reinforcement learning. We theoretically demonstrate that proper value equivalence, a method of learning models which can be used to plan optimally in the risk-neutral setting, is not sufficient to plan optimally in the risk-sensitive setting. We leverage distributional reinforcement learning to introduce two new notions of model equivalence, one which is general and can be used to plan for any risk measure, but is intractable; and a practical variation which allows one to choose which risk measures they may plan optimally for. We demonstrate how our framework can be used to augment any model-free risk-sensitive algorithm, and provide both tabular and large-scale experiments to demonstrate its ability.

Paper

References (51)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer B3wZ7/10 · confidence 4/52023-06-21

Summary

This paper addresses learning models that are sufficient to model the environment, in the sense that optimising for a risk-sensitive objective on that model is equivalent to optimising the risk-sensitive objective on the actual environment. The authors show that value-equivalent models (which match the real environment in expectation) are insufficient in the risk-sensitive case. The authors introduce the distribution equivalence principle, which defines the set of models that induce the same return distribution as the real environment (and therefore the same risk-sensitive values). To relax the assumption that the entire return distribution must be matched, the authors introduce statistical function equivalence, meaning that the models are are equivalent in terms of some statistics of the return (e.g. mean and variance). The authors define loss functions that implement these insights.

Strengths

* I think the paper addresses a novel topic: how best to learn models when the goal is to learn a risk-sensitive policy rather than a standard expected value policy. * The writing is clear in sections 2-5, and the ideas are well-formalised. I think sections 2-5 are very strong.

Weaknesses

* The paper makes it clear that value-equivalent models fail in the risk-sensitive setting (Proposition 3.2 and experiments), however it fails to motivate why one would use distribution equivalence instead of the standard approach to learning a model - maximum likelihood estimation. Presumably, if the model has enough capacity and is expressive enough, we can expect a model learnt using MLE to learn the correct distributions for distributional equivalence. The paper should explain and demonstrate in which situations the model learnt using the proposed approach results in better performance than MLE model estimate. For example, in the case where model has limited capacity, there is limited data, or the model uses a simplified distribution (such as a Gaussian over successor states), we might expect the proposed approach to work better for risk-sensitive optimisation. However, the paper does not discuss or demonstrate these potential advantages. * The description of the experiments is unclear - please see my questions. * In the empirical evaluation, it appears that the authors do not compare against MLE model estimation, which is the most obvious baseline. In Four Rooms/Frozen Lake/Windy Cliffs, the only baseline is the value equivalent model (which is obviously a bad approach for risk-sensitive optimisation). In the option-trading environment, the authors compare against the VE model again, as well as a model-free approach. The authors outperform the model-free approach, but lines 318 and 329 hint that this is because their approach is model-based (and therefore obtains better sample efficiency). Thus, it is unclear if the approach of the authors is better because of the ideas introduced in the paper, or simply because it is model-based. The authors should compare against the MLE model baseline to demonstrate that in some situations their approach is better than the most naive MLE model-based approach. * Proposition 3.2 uses this introduced notion of epsilon-strictly risk-sensitive. This definition is suitable for CVaR, but does not apply to many spectral risk measures, which apply non-zero mass to all quantiles (e.g. Wang risk measure). In the latter case, epsilon is zero, and therefore Proposition 3.2 does not support the arguments of the authors, as it shows the error is greater than or equal to zero. It seems like it should be possible to come up with a different bound for proposition 3.2 that is more general (and therefore shows that value-equivalence can be sub-optimal for any spectral risk measures other than expected value). Minor comments: * Reproducibility: the code provided has no Readme, and for the option trading environment it is completely unclear which of many files to run to generate the results in the paper. * $\Pi$ is used to denote both a set of policies, and a projection operator. This notation is a little confusing. Other than the lack of discussion about the potential advantages of the proposed approach (compared to MLE), the lack of a comparison to standard MLE baseline in the experiments, and uncertainty about some of the experiment details, I really like the paper. Thus, I am likely to increase my score to an accept score if I think these concerns are adequately addressed during the rebuttal period.

Questions

* Are you able to provide experiments demonstrating the difference in performance between between risk-sensitive policies optimised using your distribution-equivalent models vs an MLE model? Perhaps we might expect that models learnt using your approach result in better risk-sensitive performance when model capacity is limited (as demonstrated in Grimm 2020 for the value-equivalence case). * Can you explain the situations (such as limited model capacity, or limited data) where your approach is more suitable than MLE model learning for risk-sensitive policy learning? * Section 5 does a good job of explaining that models can be equivalent only for a certain set of return statistics that are of interest. However, in the tabular experiments the authors learn an equivalent model for the mean and variance, and then optimise for CVaR. This seems to contradict section 5, as mean and variance are not sufficient for estimating CVaR. What was the motivation for choosing the mean and variance as the return statistics of interest in this experiment? * What return statistics/functionals were used for the distribution-equivalent model in the option trading experiment? * Is the improvement in sample efficiency (Line 329 of the paper) compared to the Lim & Malik (2022) approach because of the distributional-equivalence approach you have proposed, or simply because any model-based approach (such as MLE) is more efficient than Lim & Malik (2022)? * How is the model represented in the experiments? In the tabular environment, I assumed it was just a categorical distribution over successor states. However, the appendix says that GPU training was used - indicating that this is a neural network model? Likewise, what was the model architecture for the option-training case?

Rating

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

Confidence

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

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

I think the authors have done a good job of addressing potential limitations. In particular, by proposing the approximate version of distributional equivalence in Section 6.

Reviewer du985/10 · confidence 3/52023-07-06

Summary

This paper studies the intersection of model-based RL and risk-sensitive RL. Firstly, the authors theoretically demonstrate that proper value equivalence can only plan optimally in the risk-neutral setting, and its performance will deteriorate as the risk level increases. Then the authors introduce distributional equivalence principle and prove that distributional equivalent models can be used for optimal planning with any risk measure. However, due to the inherent challenges of distribution, learning distributional equivalent models is not practical. Therefore, the authors combine the statistical functionals and propose statistical functional equivalence, which is parameterized by the choice of a statistical functional. The authors further demonstrate that the choice of a statistical functional determines the risk measures that can be used for optimal planning, and provide the loss functions for learning these models. Additionally, the authors show how the proposed framework can be integrated with existing model-free risk-sensitive algorithms. Finally, the authors validate the performance of the framework in both tabular experiments and option trading scenarios.

Strengths

1. The paper is well written. The original contributions are highlighted clearly. 2. This paper demonstrates clear logic and presents a series of comprehensive theoretical proofs to the validity of the proposed methods. 3. The structure of this paper is complete, it provides an illustrative example that is simple and easy to understand.

Weaknesses

1. The selection of parameters of the experimental environments and algorithms is not clearly given, such as the reward setting in Four rooms and the parameter settings of various methods in Option Trading. 2. The work in the experimental part is insufficient, and there is no more presentation of experimental results in the appendix, which causes the lack of persuasion. 3. I think the authors can add some enhanced verification experiments on the performance of more model-free risk-sensitive RL algorithms augmented with the proposed framework.

Questions

1. What is the distinction between the spectral risk measures used in the paper and the distorted expectation risk measures? 2. Why is it necessary for the weight function of spectral risk measures to satisfy a non-increasing condition? 3. Tabular experiments only show the performance comparison with respect to ($\operatorname{CVaR}(0.5)$). I think it's better to show the expected returns as well. 4. For the combination of the modification of QR-DQN and statistical functional equivalent models, it is better to provide a description of the procedures of the algorithm for understanding. 5. When using statistical functional equivalent models to augment model-free risk-sensitive RL algorithms, will the data generated by the model be added to the replay buffer to improve sample efficiency during training? 6. In Line 110, $\mu$ in equation $F_\mu^{-1}(u)=\inf \{z \in \mathbb{R}: \mu(-\infty, z] \geq u\}$ seems to have a different meaning from $\operatorname{CVaR}_\tau(\mu)=\underset{Z \sim \mu}{\mathbb{E}}\left[Z \mid Z \leq F_\mu^{-1}(\tau)\right]$. Is this true? 7. The formulas in the paper lack proper numerical labels. Please improve it. 8. In Line 126, I think there may be something wrong when you are calculating the equation $\eta^{\pi^b}(x)=U([-2,2])$, because the superposition of uniform distributions should result in a triangular distribution. Can you show me your detailed calculation steps?

Rating

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

Confidence

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

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

I think the authors need to add more discussion around the limitations of their approach. Specifically, can the proposed framework augment any model-free risk-sensitive RL algorithm? Will statistical functional equivalence limit the risk measures that can be used? A discussion on this point would be useful.

Reviewer us8H7/10 · confidence 4/52023-07-06

Summary

The authors propose to extend the notion of value equivalence to distributional model equivalence for the purpose of risk-sensitive reinforcement learning. Theoretically, the paper first shows that value equivalence is insufficient for planning risk-sensitive policies, then introduces both exact and approximate versions of distributional model equivalence, along with some of their theoretical properties. Empirically, the utility of the proposed approach is demonstrated on a number of simple domains for evaluating risk-sensitive policies.

Strengths

The main theoretical contribution of the paper is in generalizing the notion of value equivalence to distributional model equivalence for the purpose of learning risk-sensitive policies. This is well-motivated and the presentation is clear.

Weaknesses

The empirical evaluations are rather limited. It seems that sample efficiency is the key motivation for the entire approach, yet this is not clearly demonstrated by the examples in section 7. It might be more illuminating if one could see how the performance changes with respect to the number of training samples (actual and/or modeled).

Questions

What about policy gradient approaches? Is the proposed notion of distributional model equivalence still relevant?

Rating

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

Confidence

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

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer B3wZ2023-08-10

Well done on this strong rebuttal, and thank you for addressing all of the points that I raised in my review. I think the new results are great, and demonstrate that there is practical utility to this approach in additional to the theoretical contributions. I also appreciate the improvement to Proposition 3.2. In light of these improvements, I now believe that the paper should be accepted. I will update my score to a 7.

Reviewer du982023-08-15

I would like to thank the authors for providing responses to my questions. I still believe that the experimental design and results in the paper are insufficient, causing its limited persuasiveness. Therefore, I will maintain my initial rating.

Authorsrebuttal2023-08-18

Thanks! Concrete suggestion for experiments?

We appreciate your response. Thank you! Since you mentioned that the experimental design and results in the paper are insufficient, even after our new results during the rebuttal phase, we would like to kindly ask you if you have any specific experiments in mind that you would like to see? We will consider your suggestions in our future revisions, so that we have a more convincing paper.

Reviewer us8H2023-08-18

I thank the authors for the rebuttal. I'll keep my score.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC