Eliciting User Preferences for Personalized Multi-Objective Decision Making through Comparative Feedback

In classic reinforcement learning (RL) and decision making problems, policies are evaluated with respect to a scalar reward function, and all optimal policies are the same with regards to their expected return. However, many real-world problems involve balancing multiple, sometimes conflicting, objectives whose relative priority will vary according to the preferences of each user. Consequently, a policy that is optimal for one user might be sub-optimal for another. In this work, we propose a multi-objective decision making framework that accommodates different user preferences over objectives, where preferences are learned via policy comparisons. Our model consists of a Markov decision process with a vector-valued reward function, with each user having an unknown preference vector that expresses the relative importance of each objective. The goal is to efficiently compute a near-optimal policy for a given user. We consider two user feedback models. We first address the case where a user is provided with two policies and returns their preferred policy as feedback. We then move to a different user feedback model, where a user is instead provided with two small weighted sets of representative trajectories and selects the preferred one. In both cases, we suggest an algorithm that finds a nearly optimal policy for the user using a small number of comparison queries.

Paper

References (53)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer BDkG3/10 · confidence 4/52023-07-04

Summary

The paper presents a multi-objective decision making framework that accommodates different user preferences over multiple objectives. Furthermore, the framework facilitates learning these preferences via policy comparisons. The underlying model is assumed to be a finite horizon Markov decision process, while the user preference relation is assumed to be a linear combination of objective values. The preference relation is subsequently learned using user feedback taking two forms: policy comparison, and the comparison of two small sets of representative trajectories, respectively. The paper then suggests an algorithm for finding near-optimal policies using a number of comparison queries that scales quasi-linearly with the number of objectives.

Strengths

The paper addresses an important problem in decision making under uncertainty with multiple objectives which has a wide range of potential applications.

Weaknesses

However, the paper has several shortcomings. First, the presentation lacks a small running example and therefore it is quite difficult to follow the technical details of the proposed approach. Second, I was surprised by the abrupt ending of the paper (with Theorem 3) without any follow-up or discussion. It gives the impression that the submission was rushed right before the deadline which is kinda disappointing. Third, the paper lacks any empirical evidence of the proposed approach and therefore we don't have any idea whether the approach is practical or not. For example, if the MDP has a very complex state (and/or action) space, how would one best represent and present the trajectories to the user to elicit feedback. While it seems straightforward in a toy-like setting, I don't think the preference elicitation scheme based on trajectories is actually practical in a real setting.

Questions

No specific questions.

Rating

3: Reject: For instance, a paper with technical flaws, weak evaluation, inadequate reproducibility and incompletely addressed 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

2 fair

Presentation

1 poor

Contribution

2 fair

Limitations

I found the paper closely related to previous work by (Marinescu et al, 2017) which also looks at multi-objective decision making in the case of influence diagram models (which are actually more general than finite-horizon MDPs) and an unknown linear preference relation between objectives. The goal however, is slightly different, in that it is concerned with finding possibly optimal policies (which are optimal under at least one instantiation of the preference vector). (Marinescu et al, 2017) Marinescu, R., Razak, A. and Wilson, N. 'Multi-Objective Influence Diagrams with Possibly Optimal Policies', Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17), San Francisco, California, USA, 4 – 9 February, pp. 3783 - 3789.

Reviewer pTmr5/10 · confidence 4/52023-07-05

Summary

The paper presents a novel approach to elicit the unknown linear weights used to aggregate multi-objective valuations in sequential decision-making problems. The authors first consider comparative queries where a user can compare policies (instead of trajectories as in previous work). They investigate then the case where a policy can be replaced and approximated by a set of weighted trajectories.

Strengths

The paper investigates a novel approach for preference elicitation in multi-objective problems. The authors obtain upper-bounds on the number of queries needed to obtain a near optimal policy for a given user. The approach scales well with the number of users.

Weaknesses

I think the obtained results are mostly of theoretical interest, because the type of considered queries is not very practical, in my opinion. Directly comparing policies is naturally usually not possible for humans, which motivates the second case investigated in this paper. However, even comparing two sets of weighted trajectories would be very hard for humans. It is not clear if the proposed approach based on geometric and algebraic results can work in presence of noisy feedback from the user, which is a more realistic setting. The writing, while reasonably clear, seems to be a bit rushed. Notably, the paper ends abruptly, without any conclusion.

Questions

1) In footnote 1, why do the authors mention "nearly"? I think under the assumptions they make, optimal policies for every possible user preferences are in the set of Pareto optimal solutions. 2) Are there situations where comparing sets of weighted trajectories would be practical for humans? 3) Why do you not assume that the weight vector w^* is normalized since dividing by a positive value doesn't change the preferences over policies? 4) Can the proposed technique be adapted if there is noise in the user's comparisons? Some minor issues: - line 107: V(\pi) -> V^\pi - line 128: In line 96, policies are introduced as stochastic, which seems to contradict with the explicit policy representation. - line 259: vacuous -> non-vacuous?

Rating

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

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

2 fair

Contribution

2 fair

Limitations

Not applicable, this is a theoretical paper.

Reviewer 54577/10 · confidence 2/52023-07-11

Summary

The paper provides a multi-objective decision-making framework to learn approximate optimal personalized policies by eliciting user preferences via comparative queries. The authors characterize theoretically two feedback models, explicit queries and weights queries sets, and derive sound algorithms with proven upper bounds on the learning complexity.

Strengths

To the best of my knowledge, the proposed algorithms and theoretical analysis are novel and interesting. Eliciting multi-objective user preferences over policies has a tremendous impact on many (real-world) settings in which decision-making systems aid humans in daily tasks. The paper is well-written. All the theoretical assumptions, results and algorithms are well-motivated and clearly described. I checked the proofs (and skimmed through the appendix), but a more rigorous analysis would require more time.

Weaknesses

In my opinion, I can see two main areas of improvement (1) extending the algorithm to noisy feedback from the users and (2) providing some empirical evidence about the effectiveness of the proposed algorithms. Lastly, the paper is dense with information, but it ends abruptly. I would have enjoyed a paragraph either summarizing the main findings or expanding the current research to consider its potential practical impact (e.g., challenges in implementation).

Questions

- If we consider human-decision making, do you think the two proposed feedback models are realistically usable by humans or do you think they might work well with agents, for example, in cooperative AI settings?

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

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately address all the limitations.

Reviewer Gnoo7/10 · confidence 4/52023-07-18

Summary

_**Edit:** I have read the authors' rebuttal, and they have addressed all the concerns raised in my review._ The paper considers the problem of multi-objective reinforcement learning, when there is an unknown linear preference over the objectives, and approximate preference comparisons can be performed. The goal is to find an optimal policy (according to the true preference) while minimizing the number of comparisons. The main task is identifying the unknown preference, which the paper shows how to do using Algorithm 1, Algorithm 3, and Theorem 1. Once the preference is (approximately) identified, Theorem 1 shows that solving the RL problem for this approximate preference returns a policy that is competitive with the optimal policy for the true preference. Additionally, the paper also considers how preference comparisons could be implemented. One option is to ask the user to compare two policies directly, which might be hard to do. Because of this, the paper considers the case when the user can compare trajectories from two policies, which might be more interpretable to the user. To this end, the paper shows how to efficiently find a small (O(k)) weighted set of trajectories such that their weighted values are equal to the value of the policy (Algorithm 2).

Strengths

- The procedure for identification of linear preference is not limited to just RL and can be used in a wider setting, making it of general interest. - Algorithm 1 adapts to $d \le k$, which could be helpful in cases when there are many redundant objectives. - The weighted trajectory set representation of a policy is novel, interpretable, and possibly of wider interest.

Weaknesses

- Line 220: “we only need to run Algorithm 1 once for all users”. This seems to be incorrect, as Algorithm 1 uses preference comparisons which depend on the user. This means the leading factor in the bound on line 256 should change from $(k^2+n)$ to $n (k^2 + 1)$. - The paper assumes the existence of a “do nothing” action, which might require a modified MDP. While this appears straightforward, it seems to me it requires domain-specific knowledge in order to modify it such that the “do nothing” policy has minimal value (which is the true requirement). E.g. in the chess example, perhaps refusing to play is considered a draw, which is better than a loss. Thus, a “do nothing” action would need to somehow always lose.

Questions

- Proof of Lemma 1: in the first inequality, shouldn't it be possible to reduce $k \epsilon$ to $\epsilon$? This would make Lemma 1 applicable to when $\epsilon \le v^\*/(2k)$ rather than $v^*/(2k^2)$. - Is it possible to approximately estimate $\epsilon$ by comparing $\pi^1$ against a scaled version of itself (similar to the binary search in Algorithm 3)? If $\hat{\epsilon} \approx \epsilon$ can be estimated this way, can Theorem 2 be used with $\delta = k^{5/3} \hat{\epsilon}^{1/3}$, avoiding the problem of having to guess the hyperparameter $\delta$? - Have you considered the case when the user can compare policy value vectors? In this case, we can query with “artificial” values directly (e.g. basis vectors), requiring no interaction with the MDP at all. Perhaps this makes the problem too simple? - The preference identification procedure can be extended to general, non-tabular RL case. Is the same true for efficiently finding weighted trajectory sets?

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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

The paper requires that preferences are linear, which is a fairly strong assumption. However, it is a good first step.

Reviewer wfQV5/10 · confidence 3/52023-07-19

Summary

This paper proposes a method for multi-objective RL considering two user feedback models and thereby the first algorithms with theoretical guarantees for personalized MO decision making via policy comparison.

Strengths

The paper prpoposes an interesting method for finding a personalized optimal policy for MORL problems reflecting users' unknown preferences over the objectives. It is very well written and scienitifically sound, motivates the work well. Method is presented clearly and carefully. The novelty and significance of the method are well presented.

Weaknesses

The paper lacks conclusions and discusssion about limitations completely and thereby its utility is left a bit open. The paper also emphasizes that it is important to present the obtained trajectories in an easy format to the user, but does not give any examples how this would be achieved.

Questions

You state that the method learns a user's preference vector using minimal number of queries, and C4 is used when the number is small, but discussion of what is small in the number of trajectories or minimal for queries? Could you give an intuitive example of how are the easily interpretable queries presented to the user. What are the limitations of the method?

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

4 excellent

Contribution

3 good

Limitations

Not really addressed.

Reviewer Gnoo2023-08-10

Thank you for the reply. All my concerns have been addressed, except for the (very minor) one regarding Lemma 1: > We agree with the reviewer that we are able to reduce $k \epsilon$ to $\epsilon$ if we compare every pair of $\pi^{e_i}$ and $\pi^{e_j}$. However, our prime focus is on minimizing the number of queries and this will increase the number of queries (used for searching $\pi_1$) from $k$ to $k^2$. Hence, it is a trade off between the number of queries and accuracy. We'll be sure to address that in the final version of the paper. I believe the $O(k)$ approach in the paper is enough to get $\epsilon$ optimality. Here is an attempt at a proof, by induction. In the base case, comparing $k = 2$ policies, it is clear the returned policy will be $\epsilon$-optimal. For $k > 2$, first run the comparison on the first $k-1$ policies to get an $\epsilon$-optimal policy $\pi^*$: for all $i \in [k-1]$, $w \pi^* \ge w \pi_i - \epsilon$. Then, compare $\pi_k$ and $\pi^*$. Two cases: - If $w \pi^* < w \pi_k - \epsilon$, then $\pi_k \succ \pi^*$, so $\pi_k$ will be returned, and also $w \pi_k \ge w \pi_i - \epsilon$ for all $i \in [k]$, so $\pi_k$ is $\epsilon$-optimal over all $k$ policies. - Otherwise, $w \pi^* \ge w \pi_k - \epsilon$, in which case $\pi^*$ is returned, we have $w \pi^* \ge w \pi_i - \epsilon$ for all $i \in [k]$, so $\pi^*$ is $\epsilon$-optimal over all $k$ policies.

Authorsrebuttal2023-08-11

Replying to \epsilon optimality

We thank the reviewer for reading our rebuttal and responding promptly. We agree with the reviewer that $O(k)$ is enough to guarantee $\epsilon$ optimality by the induction method mentioned by the reviewer! We will incorporate this into the final version and update the results.

Reviewer 54572023-08-11

I thank the authors for the rebuttal! At the present time, I do not have additional questions, since the various answers cleared most of my doubts. Lastly, I stand with my previous scores.

Reviewer pTmr2023-08-16

Rebuttal

Thank you for the rebuttal. Regarding the set of weighted trajectories, in all your examples, the trajectories are summarized by some scalar values. However, in the multiobjective setting, comparisons become much harder in my opinion. That was what I meant. Regarding the discussion at the end of the paper and the addition of a running example, it's not clear how the authors would manage that, given the limited space they already have. For now, since I don't see any updated submission, I will keep my score unchanged.

Authorsrebuttal2023-08-17

Thank you for reading our rebuttal and quick response. We will include this discussion in the additional content page that is allowed for accepted papers (https://neurips.cc/Conferences/2023/CallForPapers). Regarding the trajectory sets- we apologize for the misunderstanding. Here are extensions of two of our examples for multiple objectives: Investment strategy- when considering investing money in a specific portfolio (policy), clear objectives are: 1. Expected reward. 2. Certain risk measures that are non-linear in the reward (e.g., VaR, CVaR). 3. The amount of money that the investor might lose. For example, in short selling, one could lose more money than they initially invested. 4. The Eco-friendliness of the portfolio throughout time. For example, investing in big oil companies might be less eco-friendly compared to, e.g., Seventh Generation (a cleaning product company that sells eco-friendly products). Notice that while portfolios could have the same expected reward, their other criteria could vary significantly. One example of the weighted trajectory set of an investment strategy is: 1. 90% weight on a trajectory of reward 7, low-risk measures in all states, huge losses in a few states, and very eco-friendly investments; 2. 10% weight on a trajectory of reward -6, high-risk measures in half of the states and low-risk measures in the other half, huge losses in some states, and with the least eco-friendly investments. Commuting to the airport- besides commute time, one could consider potential productivity (as people might be more productive when they are on the train), comfort, safety, and again eco-friendliness. We would be happy to incorporate these examples in the paper!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC