An Offline Adaptation Framework for Constrained Multi-Objective Reinforcement Learning

In recent years, significant progress has been made in multi-objective reinforcement learning (RL) research, which aims to balance multiple objectives by incorporating preferences for each objective. In most existing studies, specific preferences must be provided during deployment to indicate the desired policies explicitly. However, designing these preferences depends heavily on human prior knowledge, which is typically obtained through extensive observation of high-performing demonstrations with expected behaviors. In this work, we propose a simple yet effective offline adaptation framework for multi-objective RL problems without assuming handcrafted target preferences, but only given several demonstrations to implicitly indicate the preferences of expected policies. Additionally, we demonstrate that our framework can naturally be extended to meet constraints on safety-critical objectives by utilizing safe demonstrations, even when the safety thresholds are unknown. Empirical results on offline multi-objective and safe tasks demonstrate the capability of our framework to infer policies that align with real preferences while meeting the constraints implied by the provided demonstrations.

Paper

Similar papers

Peer review

Reviewer TEUx6/10 · confidence 4/52024-06-26

Summary

This paper presents a novel framework for multi-objective RL that eliminates the need for handcrafted target preferences. Instead, it uses demonstrations to implicitly indicate the preferences of expected policies. The proposed offline adaptation framework can also handle safety-critical objectives by utilizing safe demonstrations, even when safety thresholds are unknown. The authors provide empirical results showing that the framework effectively infers policies that align with real preferences and meet the constraints implied by the demonstrations.

Strengths

(1) Interesting topics: Developing a framework for multi-objective reinforcement learning that does not rely on handcrafted target preferences is very interesting and potentially useful for real-world applications. Additionally, the framework uses demonstrations to implicitly indicate the preferences of expected policies, which can also be extended to meet constraints on safety-critical objectives using safe demonstrations. (2) Extensive experiment evaluation: The authors provided extensive experiment results including those on offline safe RL benchmark, offline RL benchmark, and multi-objective RL benchmark. These results show the flexibility and effectiveness of the proposed method.

Weaknesses

(1) Preference is closely related to the reward function definition. How does this definition compare to other metrics such as the ones based on the states/state-action/latent variable distribution? For some tasks, there exists some correlation between reward functions, and this will potentially affect the preference estimation results. For other potential weaknesses, please see my questions.

Questions

(1) Preference is closely related to the reward function definition, and correlation among reward functions will result in preference calculation issues. For example, what if $r_1 = 0.5 r_2 = 0.2 r_3$, and $r_4$ is independent to $r_1$? In this setting, we would get a small preference mismatch score between two demonstrations even if they have significant differences in the reward returns for $r_4$. Do you have any methods to deal with this? (2) In section 4.2, the first step of your method is to learn a set of policies that respond to various preferences during training. How do you ensure that the learned policy is subject to specific preferences? (3) Based on the description in section 4.3, are High-reward-high-cost and low-reward-low-cost demonstrations categorized as similar behaviors? (4) How does the proposed method perform in other DSRL datasets/tasks such as Car-Circle and Drone-Circle? (5) Where are CDT results discussed in section 5.2? I did not them in Figures 3 and 4. (6) How do your method and problem formulation relate to multi-constraint safe RL since multiple constraints can also be viewed as multiple objectives? Some related works: [1] Kim, D., Lee, K., & Oh, S. (2024). Trust region-based safe distributional reinforcement learning for multiple constraints. Advances in neural information processing systems, 36. [2] Yao, Y., Liu, Z., Cen, Z., Huang, P., Zhang, T., Yu, W., & Zhao, D. (2023). Gradient Shaping for Multi-Constraint Safe Reinforcement Learning. arXiv preprint arXiv:2312.15127. [3] Guan, J., Shen, L., Zhou, A., Li, L., Hu, H., He, X., ... & Jiang, C. (2024). POCE: Primal Policy Optimization with Conservative Estimation for Multi-constraint Offline Reinforcement Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 26243-26253).

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, the authors addressed the limitations.

Authorsrebuttal2024-08-10

Thanks again for all the insightful comments and suggestions. They are helpful in improving our paper.

Reviewer 58tv6/10 · confidence 3/52024-07-17

Summary

The paper presents a new multi-objective RL setup to learn a policy offline that can adapt to unknown target preference online, utilizing posterior estimates of the true target preference.

Strengths

1. The paper is very clearly written. 2. The base policy is reasonable (decision transformer and diffusion policy). 3. Experiment and figures are easy to follow. 4. The methodology is built on solid prior work and the extensions are easy to understand.

Weaknesses

After carefully thinking through this setup, I'm a bit worried about the core contribution of this paper. There are several major challenges to offline learning of CMO-MDP: 1. Policy learning conditioned on preference vector (solved by previous works and used in this paper) (MODF, MORvS) 2. Trajectory distributions shift under the optimal policy and approximate policy, leading to failures of estimating preference vector w (solved by [Ghosh et al., Hong et al., 2022], Eq 6, used in this paper) 3. Preference vector online estimation (through posterior update) (Eq 5) (this is a straightforward application), and the actual update is Eq 7. It seems to me that the paper's core contribution is: 1. Proposed a new problem. 2. Proposed a good preference target vector estimation (Eq 7). 3. Cast constrained problem as unconstrained and added a conservative estimate to the preference vector so that constraints are not violated during adaptation. I'm not fully able to determine how significant these contributions are.

Questions

For Eq 5, can the author comment on -- what is the estimation quality of the target preference policy if the sampling policy pi-hat-star-w is sub-optimal? I can break this question down to two parts: - How does the sub-optimality of the approximate policy to optimal policy affect the estimation of w? - How does the estimation quality of w affect the convergence of the approximate policy to the true optimal policy?

Rating

6

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

The authors addressed the limitation.

Reviewer ZqLH6/10 · confidence 3/52024-07-18

Summary

This paper addresses the issue of handcrafted target preferences in multi-objective RL (MORL) by proposing an offline adaptation framework called Preference Distribution Offline Adaptation (PDOA), which implicitly learns the preferences from a small number of demonstrations. Specifically, PDOA involves two steps: (i) It learns a set of policies for various preferences during training using an off-the-shelf MORAL algorithm; (ii) It then adapts a distribution of target preferences based on the demonstrations at deployment. To update the posterior of target preferences in a tractable manner, PDOA adopts multiple approximation schemes for both unconstrained and constrained settings. Moreover, in the constrained setting, to address the discrepancy between the true preference and the estimated one, PDOA also proposes to use a conservative estimate with the help of CVaR. The PDOA is evaluated on multiple MOD4RL locomotion tasks and has strong performance in terms of average utility and hypervolume. Empirical studies on the number of demonstrations and the conservatism parameter are also provided.

Strengths

- This paper identifies and addresses the issue of target preference estimation by offline preference adaptation, which is a novel setting in the literature of MORL. - The idea of rethinking target preferences from an Bayesian inference perspective is quite interesting. The overall approach looks reasonable (despite some concerns listed below). - The experiments are done on multiple D4MORL tasks, and the results in terms of utility and hypervolume are quite promising. - The ablation studies are quite helpful in understanding the robustness of PDOA.

Weaknesses

- One issue is the lack of stronger offline imitation learning baselines. If we already have a number of demonstrations of the target preferences at test time, then one reasonable solution is to view this as a single-objective problem and then apply an off-the-shelf offline imitation learning (IL) algorithm with these demonstrations, such as DWBC [Xu et al., 2022] and DemoDICE [Kim et al., 2022], which both focus on offline imitation with a small number of expert demonstrations along with other sub-optimal trajectories. In the setting of PDOA, one can simply do offline IL by using $\mathcal{B}_\omega$ as the expert demonstrations and $\mathcal{D}$ as the sub-optimal samples. - Another concern is on approximation of the posterior in (6). While I can get the intuition that in-distribution samples tend to have higher $r^{TD}$, the surrogate in (6) appears to be quite heuristic and not very principled. Specifically, the $\pi^*$ in the original posterior expression could be different from the policy (whose Q and V are used in Equation 6) learned by MORL at the first stage. More theoretical justification of this approximation is needed. - For the constrained setting, there is an additional parameter $\alpha$ to determine to strike a balance between constraint violation and achieved utility. However, my concern is that $\alpha$ is a parameter to be used at test time, and hence there is very little one can do in terms of parameter tuning. While Figure 6 did provide a study on the influence of $\alpha$, we can see that this value does matter a lot in the balance between constraint violation and utility. As a result, this could make PDOA less attractive in practice. [Xu et al., 2022] Haoran Xu, Xianyuan Zhan, Honglei Yin, and Huiling Qin, “Discriminator-Weighted Offline Imitation Learning from Suboptimal Demonstrations,” ICML 2022. [Kim et al., 2022] Geon-Hyeong Kim, Seokin Seo, Jongmin Lee, Wonseok Jeon, HyeongJoo Hwang, Hongseok Yang, and Kee-Eung Kim, “DemoDICE: Offline Imitation Learning with Supplementary Imperfect Demonstrations,” ICLR 2022.

Questions

Please see the comments above. Some additional questions: - How many trajectories (in the demonstration set) are needed for target preference adaptation in the main experiments (Figures 1-2)? - In Equation 5, during the preference estimation phase, the dataset $\mathcal{D}$ only takes part in the prior $P(\omega| \mathcal{D})$. Under PDOA, it could be possible to leverage the information from $\mathcal{D}$ to further improve the accuracy of the target preference. Can the authors comment on this?

Rating

6

Confidence

3

Soundness

2

Presentation

3

Contribution

3

Limitations

The paper discusses one limitation about the need for human expertise in constructing demonstrations in Section 6.

Reviewer ZqLH2024-08-12

Thanks for the response

Thank the authors for the detailed response. My concerns about the baselines and the demonstration set have been nicely addressed. Some follow-up comments / questions: - About W2: My question about the justification for $r^{TD}$ remains unresolved. To justify the use of (6), one would need to roughly show that $r^{TD}$ is generally larger when the posterior is larger. Somehow I do not see why the explanation in the rebuttal is relevant. Please let me know if I missed anything. - About W3: It is mentioned that “...a roughly designed $\alpha$ will not significantly undermine our approach.” However, in Figure 6(c)-(d), it appears that a difference of 0.2 in $\alpha$ can indeed have a significant impact on both the cost and the utility. Moreover, as the ablation study on $\alpha$ is only conducted on two tasks (and not evaluated extensively), the influence of this parameter can possibly be even more substantial. While I can see tuning $\alpha$ might be possible in some cases and understand what the authors try to say by ``Figure 6 demonstrates that our approach maintains relatively few constraint violations even when the conservatism mechanism is removed (i.e., $\alpha=1$),” I do feel that the statement can be an over-claim.

Authorsrebuttal2024-08-13

Apologies for the confusion. We would like to supplement the related explanations about W2 and W3. **W2: Justification of TD-reward surrogate** We will use an example to illustrate the effectiveness of the TD-error approximation. Let's assume there are three types of transitions: $(s,a,r_1,s_1),(s,a,r_2,s_2),(s,a,r_3,s_3)$, with the proportions $p$, $1-p$ and $0$ in the dataset, respectively. For the out-of-distribution transition $(s,a,r_3,s_3)$, it is expected to associate with high uncertainty and a large TD-error, according to previous research [1,2,3] and our empirical results in Figure 14. For the in-distribution transitions $(s,a,r_1,s_1),(s,a,r_2,s_2)$, the Q-value can be expressed as $Q(s,a)=p\cdot (r_1+V(s_1))+(1-p)\cdot(r_2+V(s_2))$. Thus, $(s,a,r_1,s_1)$ corresponds to a low TD-error $||Q(s,a)-r_1-V(s_1)||^2_2$ when it has a high probability of occurring in the dataset (i.e., when $p$ is close to $1$). Therefore, we expect to obtain $TD(s,a,r_1,s_1)<TD(s,a,r_2,s_2)<TD(s,a,r_3,s_3)$ when $p$ is high. In this case, TD reward aligns with the estimated dynamics $\hat P(s',r'|s,a)$ in the posterior for in-distribution transitions, where a higher occurrence probability leads to a higher TD reward. For out-of-distribution transitions, while the TD reward might not align with the estimated dynamics, it remains low to mitigate the adverse effects of inaccuracies in dynamics estimation. Similarly, for state distribution, a higher occurrence probability of states corresponds to a more accurate value function estimate and a higher TD reward, whereas out-of-distribution states typically result in a smaller TD reward. I hope this example provides an intuitive insight into this approximation. Rigorous theoretical derivations would involve the generalization of Q-value neural networks and their relationship to dynamics uncertainty, which are relatively complex and challenging. We would leave further theoretical analysis to our future work. Another potential direction is to adopt alternative approaches, such as explicit dynamic modeling, which could offer more straightforward theoretical results. [1] Bai, Chenjia, et al. "Pessimistic Bootstrapping for Uncertainty-Driven Offline Reinforcement Learning." ICLR 2021. [2] Rigter, Marc, Bruno Lacerda, and Nick Hawes. "One risk to rule them all: A risk-sensitive perspective on model-based offline reinforcement learning." NIPS 2024. [3] Ghosh, D., Ajay, A., Agrawal, P., & Levine, S.. Offline rl policies should be trained to be adaptive. ICML 2022. **W3: Conservative parameter $\alpha$** Our statement that "a roughly designed $\alpha$ will not significantly undermine our approach" was based on comparisons with the baselines, where our approach showed relatively fewer violations and a higher correlation between cost and reward (indicating the high policy diversity we expected) than baselines under various $\alpha$. We know this isn't a direct comparation shown in Figure 6 and apologize for any confusion caused by some of our claims in the response. We acknowledge that the influence of this parameter on the final performance cannot be overlooked, and we appreciate the reviewer for highlighting this. We believe that fine-tuning this parameter using a limited iteration of online interaction and re-adaptation could indeed address this issue. Additionally, the tasks used in the ablation experiments were randomly selected, and we plan to conduct these experiments across a broader set of tasks and report the results in a future version. Thank you again for the constructive feedback. We are happy to provide further explanations if there are still any confusions.

Reviewer FywX7/10 · confidence 3/52024-07-29

Summary

The paper tackles the important problem of adapting multi-objective policies to diverse preferences and constraints at deployment. This work uses state-of-the-art multi-objective RL methods to learn a set of policies that can align with diverse preferences. At deployment time, it doesn't assume access to the explicit hand-crafted preferences but introduces an adaptation method that uses demos to infer the target preferences. Additionally, the authors formulate this method to take into account explicit safety constraints at deployment time, inferred using only the demonstrations, without any online interaction. The approach is tested extensively on standard multi-objective and safe RL benchmarks, and they show that it can achieve higher performance, incur lower costs, and generate more diverse policies in these environments.

Strengths

1. The paper tackles an important problem of adaptation in multi-objective and safe RL. The intuition of the overall method is well understood, the framework and the parameters are mentioned, and the results are shown over the standard environments. 2. The paper is well organized and presents a clear background of the past work in multi-objective RL. However, it would be nice to see a clearer distinction between the past approaches (PEDA etc.) used and changes (if any) made by the authors in this paper, for phase 1 of the methodology. 3. Past work assumed access to hand-crafted preferences or online interactions to infer the preferences at deployment time. However, this work bypasses that by using a Bayesian framework to infer the preference distribution at deployment time using a set of demos explaining the preferred behavior. This is a more practical and scalable approach to adapting policies at test time. 4. This method introduces an adaptation framework to infer the distribution of preferences from demonstrations. The authors lay out the theoretical and practical details for this step, including replacing the intractable Bayesian inference with a TD-error-based approximation from past literature. 5. The introduced method is generalized to include safety constraints in the RL agent. This transforms the constrained RL problem, into an unconstrained multi-objective RL setting, which is compatible with the methods introduced in this work. Further, to ensure that the given approximation doesn't lead to excessive safety violations the authors use a CVaR-based objective, to always overestimate the constraints in case of uncertainty due to low demos. 6. The authors demonstrate the effectiveness of their approach in several multi-objective and constrained simulation environments. The paper provides ablations to compare the effect of some hyperparameters and training methods on the efficacy of the method.

Weaknesses

1. At training time, the method assumes access to the explicit preferences and demonstrations responding to those preferences. However, in practical settings, this information is not available even at training time, and the preferences and their clusters have to be learned in an unsupervised / self-supervised setting. I would be curious to know what the authors think about the effectiveness of this approach to learn the preferences and the policies in an end2end framework. 2. I would recommend the authors to present a clear distinction between the past work, and the approach finally used in this work to generate the multi-objective policies. In this current presentation, it is slightly confusing to understand the particular contributions/adaptations of past work to incorporate this with the posterior approximation framework. A pseudocode explaining the entire algorithm would make the method clearer to the readers. 3. In the background, it would be nice to see some discussion contrasting the work in multi-objective adaptation to inverse RL, because in the introduction the authors mention - “it is more natural to infer expected behaviors through a few demos that implicitly indicate….” Comparing this to the approaches in inverse RL, and further, in Bayesian RL would be very useful to highlight the contributions. 4. Under expert demos, the prompt-MODT baseline performs comparably to the methods introduced in this work. I would recommend the authors provide a wider discussion of the effectiveness of explicitly modeling human preferences using the adaptation module and if using the distributional Bayesian approach makes it robust to noisy samples.

Questions

1. The method is based on settings where the total reward is a linear combination of preferences and individual rewards. How can this setting be adapted to settings where we have access only to the total reward and demos and not the individual components? 2. In the evaluation metrics, I would be curious to know why is the hypervolume and therefore, the diversity in the policies important. For the baselines, we observe that the policies with low hypervolume still have considerably close performance to this work in terms of utility and cost. A paragraph along these lines would allow a broader audience to understand the significance of these experiments and results. 3. In certain environments, despite violating constraints, the returns are high, which indicates that the cost and the rewards are independent of each other. This is slightly confusing and I would appreciate some clarification in this respect. 4. In Figure 6, it would be nice to analyze why higher demos incur a higher cost.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

1. The method assumes access to the preferences at training time, which might not be practical for real-world settings, where the individual objectives and the preference distribution are unknown. This could be the reason that the method outperforms the prompt-MODT baseline i.e. the error introduced due to implicitly modeling the preferences, and not having access to the target distribution could potentially decrease the effectiveness of PDOA. 2. The authors mention that the adaptation step is not real-time, which makes it difficult to transfer this system to real-world RL systems such as LLMs and robots. 3. The evaluation is limited to locomotion environments, so, I suggest the authors include a discussion on the implications of the method and evaluation towards adaptation and safety in real-world RL problems.

Reviewer TEUx2024-08-09

Thanks for the additional explanation. All my concerns and questions have been successfully addressed. I have raised my score from 5 to 6 in favor of acceptance.

Reviewer FywX2024-08-11

I thank the authors for their response. I have read the rebuttal and maintain my score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC