Honor Among Bandits: No-Regret Learning for Online Fair Division

We consider the problem of online fair division of indivisible goods to players when there are a finite number of types of goods and player values are drawn from distributions with unknown means. Our goal is to maximize social welfare subject to allocating the goods fairly in expectation. When a player's value for an item is unknown at the time of allocation, we show that this problem reduces to a variant of (stochastic) multi-armed bandits, where there exists an arm for each player's value for each type of good. At each time step, we choose a distribution over arms which determines how the next item is allocated. We consider two sets of fairness constraints for this problem: envy-freeness in expectation and proportionality in expectation. Our main result is the design of an explore-then-commit algorithm that achieves $\tilde{O}(T^{2/3})$ regret while maintaining either fairness constraint. This result relies on unique properties fundamental to fair-division constraints that allow faster rates of learning, despite the restricted action space. We also prove a lower bound of $\tilde{\Omega}(T^{2/3})$ regret for our setting, showing that our results are tight.

Paper

Similar papers

Peer review

Reviewer V1uC7/10 · confidence 4/52024-07-01

Summary

The paper studies online fair division of indivisible items, where items arrive one at a time, and each item must be permanently assigned to a player upon arrival. Player utilities are initially unknown and must be learned via bandit-style feedback: there are a finite number of types of goods, and each player’s value for a good of type k is drawn independently from a distribution D_k. The authors aim to maximize utilitarian welfare (the sum of utilities) while satisfying a hard constraint of either envy-freeness in expectation (EFE) or proportionality in expectation (PE). EFE and PE are evaluated with respect to the randomized allocation on each time step. That is, before observing the type of the new item, the learner proposes a randomized allocation X_t which determines how the item will be allocated for each possible type. Each X_t is required to satisfy the fairness constraint (EFE and PE), not the cumulative realized allocation. I initially found this rather unintuitive, but the authors effectively justify this choice by showing that: 1. Any EFE algorithm achieves realized cumulative envy at most $\sqrt{T} \log(T)$, while optimal is at least $\sqrt{T}$. 2. EFE at each time step does not cause any welfare loss compared to only requiring EFE at the end. The main result is a simple explore-then-commit algorithm which is EFE or PE while achieving regret $\tilde{O}(T^{2/3})$ wrt welfare. Along the way, the authors prove that the EFE and PE constraints satisfy the following property: either all players receive the same allocation, or the constraint can be satisfied with slack. This novel result formalizes the intuition that the “hardest case” for fair division is when all agents have the same utilities. EFE/PE can trivially be satisfied in that case by the uniform allocation, and in all other cases, EFE/PE can be satisfied with slack.

Strengths

The presentation is clear and accessible. Although some of the modeling choices felt initially unintuitive, the authors did an excellent job of justifying those choices, as discussed above. The main result is satisfying, both in terms of the simple explore-then-commit algorithm and the regret bound. Essentially, the authors successfully solved the problem they chose to study. I found the either-uniform-allocation-or-slack property (and the result that EFE and PE satisfy it) to be quite interesting as well. This result captures an important intuition in a simple way. The authors argue that this property may be of independent interest, and I agree.

Weaknesses

Although I think the problem is reasonable and meshes well with prior work, I question whether the mathematical model has any real-world relevance. The authors use a running example of a food bank with multiple branches, with the idea that unpredictable perishable donations must be allocated immediately and fairly among the branches. I have several problems with this analogy: 1. Surely food banks wish to allocate according to need without worrying that one branch might “envy” another if the second branch simply has higher need? 2. The paper assumes additive utility, and each branch’s utility would be extremely non-additive (no one needs infinite cereal). 3. It’s hard to imagine the donations being so unpredictable. To be clear: I’m not claiming that a mathematical model must perfectly match reality in order to provide real-world insights. But this paper’s model seems only tenuously connected to the real-world scenario that is used as motivation. If the authors have further justification for the food bank analogy, I would be interested to hear it. To me, the most compelling story in this paper actually centers on the either-uniform-allocation-or-slack property. The authors establish this intuitive but nontrivial property, and demonstrate how it can be applied in a bandit context. I can easily imagine this property being useful for other fair division problems that might have more real-world relevance. This is more minor, but I think the citation below is quite relevant and should be discussed along with Benadè et al. Hakuei Yamada, Junpei Komiyama, Kenshi Abe, and Atsushi Iwasaki. Learning fair division from bandit feedback. International Conference on Artificial Intelligence and Statistics, 2024.

Questions

Do you agree with my criticisms of the food bank analogy? Is there a different real-world situation that you believe is captured by your mathematical model? More broadly, do you believe that your mathematical model provides real-world insights, and if so, what are they? Below is feedback that you're welcome to take or leave. I don't expect responses, even though some are phrased as questions. 1. I would suggest naming Property 2 so that I don’t have to keep calling it the either-uniform-allocation-or-slack property :) 2. “This distribution is independent of both player i’s distributions for other item types and other players’ distributions.” Distributions can’t really be independent; I think this should state that samples from those distributions are independent. 3. “Intuitively, this is equivalent to assuming that players have non-zero values for all item types and that their values have a finite upper bound.” I don’t think this is quite right. Assuming non-zero values would just require mu_{ik}^* > 0. I would describe this assumption as saying that the distribution means are “well-conditioned”. 4. Definitions 1 and 2: mu_i^* is a vector, right? I don’t think this was ever explicitly defined. (Only mu^* and mu_{ik}^* were, I think.) The definitions are also missing “for all i”. 5. “proof” should be capitalized at the start of proofs 6. Why not just use gamma_0 in Property 2? Why have for all gamma < gamma_0? 7. Is it intentional for the “2” in C_{P2} to be a hyperlink to property 2? 8. Consider mentioning the actual value of (or a bound on) gamma_0 is for each property? 9. For readability in Algorithm 1, consider writing indicator variables as \mathbb{1}(stuff) instead of \mathbb{1}_{stuff}. 10. The notation “\mu in \hat{\mu} \pm \varepsilon” is non-standard to my knowledge. Consider “\mu in [\hat{\mu} - \varepsilon, \hat{\mu} + \varepsilon]” 11. I found lines 154-178 to be quite compelling and and also quite important. Consider making this its own subsection? And maybe even including some of the Theorem statements from Appendix B? 12. I really like lines 323-330.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

The authors thoroughly discussed the technical limitations of their work, but did not discuss what I see as the greatest limitation: the weak motivation of the problem and the potential lack of real-world insights.

Reviewer MmvH8/10 · confidence 3/52024-07-13

Summary

The authors study an online fair division problem where items arrive online and need to be assigned to agents to maximize welfare while satisfying one of envy-freeness or proportionality. The novel consideration in their model is that the item valuations are drawn from an unknown distribution. They approach this as both a learning problem as well as an allocation problem. They provide a simple algorithm with small regret compared to an algorithm that knew the distributions beforehand. They also prove several structural results regarding their model.

Strengths

The model is quite interesting, as are their results. The proofs are involved, and might be of independent interest.

Weaknesses

N/A

Questions

Is it possible to achieve non-trivial regret in the setting where the valuations of the agents are not independent? In the motivating example of the food pantry, if a delivered food item is of bad quality, the agents might all lower their valuation of this item together.

Rating

8

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

N/A

Reviewer Mvzr7/10 · confidence 4/52024-07-26

Summary

In this work, the authors consider the problem of maximising the total utility in online fair allocation settings, where: (i) T items having types in set [m] arrive online in T distinct rounds; (ii) the value $V_{i}(t)$ that each agent $i$ assigns to each item $t\in [T]$, when its type is $k$, is independently drawn from a sub-Gaussian distribution with unknown mean $\mu_{i,k}^*$; (iii) for each item that arrives online, its type is known, and based on this information and previous observations, it must be irrevocably assigned to some agent; (iv) the value $V_{i}(t)$ of each item $t$ can only be observed after it has been assigned, and this information allows the estimation of the unknown probability distributions as $t = 1, \ldots, T$ varies. The authors design an Explore-Then-Commit algorithm that maximises the total utility of the various players over $T$ rounds, with a regret of the order $\tilde{O}(T^{2/3})$, where the maximisation is constrained by achieving fairness properties (envy-freeness or proportionality) in expected value, with high probability.

Strengths

-This paper introduces an interesting setting of online fair allocation, where techniques derived from bandit algorithms are applied to the case where valuations are distributed according to unknown distributions. -The approaches considered in Lemmas 1 and 2, which characterise fairness properties robust to small uncertainties in the underlying distribution, are interesting and novel. -The paper certainly required a considerable technical effort.

Weaknesses

I believe that the results could be written and presented more effectively (see the major comments below). Another minor criticism is that, although the results provided in Lemmas 1 and 2 are quite novel and there has been a considerable technical effort in writing the complete version, the results obtained are somewhat expected within the field of bandit algorithms and follow standard approaches. Major comments: -It should be highlighted from the beginning that the allocations returned by the algorithm are envy-free in expectation, but with high probability (that is, not always). -The matrix notation could sometimes be avoided in favour of explicitly listing the various constraints. -It would be clearer to present the main algorithm right from the start and explain informally how the various properties come into play. -The fact that the presented LP has infinite constraints is mentioned in the conclusion. I suggest providing more details in the technical section. Minor comments: -The acronyms efe and pe are not defined (despite clear from the context). -Remark 1: specify that the rows different from i and i' are composed of zeros only. -Claims of Lemma 1 and Lemma 2: satisfy -> satisfies.

Questions

No questions.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer V1uC2024-08-10

I mostly find the authors' response compelling, and I will raise my score by 1 point under the assumption that the further justification for their mathematical model is incorporated into the final version of the paper. The most compelling point in my opinion is the existence of a real application that matches this mathematical model, and I would recommend the authors discuss Food Drop explicitly in the final version. I will briefly mention that I find this aspect of their response unconvincing > Additive utility is a standard assumption for online fair division due to the difficulty of studying non-linear utilities The fact that the assumption is standard doesn't make it realistic. And the fact that non-linear utilities are challenging to analyze doesn't make the actual utilities additive.

Authorsrebuttal2024-08-10

Your response is greatly appreciated. We're happy to follow your suggestion and discuss Food Drop. > The fact that the assumption is standard doesn't make it realistic. And the fact that non-linear utilities are challenging to analyze doesn't make the actual utilities additive. This is well taken. Not to belabor the point, but perhaps the argument we briefly mentioned about the frequency of allocations is more compelling. For example, in the food allocation data of Lee et al. (2019), there were a total of 1760 donations from 169 donors over the course of five months, and 277 organizations received donations. This means that organizations receive donations every 3-4 weeks on average, suggesting that the donations can largely be regarded as independent. Min Kyung Lee, Daniel Kusbit, Anson Kahng, Ji Tae Kim, Xinran Yuan, Allissa Chan, Daniel See, Ritesh Noothigattu, Siheon Lee, Alexandros Psomas, Ariel D. Procaccia: WeBuildAI: Participatory Framework for Algorithmic Governance. Proc. ACM Hum. Comput. Interact. 3(CSCW): 181:1-181:35 (2019)

Reviewer V1uC2024-08-10

I think the frequency argument is reasonable. More generally, I think some empirical work may be warranted to determine the extent to which utilities are in fact additive, since as you say, this is quite a common assumption in this area. I understand that that's beyond the scope of the current submission, of course.

Reviewer Mvzr2024-08-12

Rebuttal

I have read the authors' response carefully, and I am satisfied with their reply. I strongly believe that the rationale behind the considered setting will be even more exhaustive in the revised version. Moreover, although the case of additive functions is more restrictive compared to general monotone functions, I do not think this is an issue, as this setting is one of the most studied in the field and is well established in the literature. However, I believe that the interest in this setting should not be justified by the fact that the more general setting is difficult or not possible in online learning. Finally, I think that the tight lower bound presented by the authors deserves to be mentioned in the new version (including the complete proof in the appendix), as it further justifies the approach obtained for the upper bound. In light of the above comments, I will increase my score by one point.

Authorsrebuttal2024-08-12

Thank you very much for your comment. We appreciate your take-aways from the entire discussion and are happy to follow your suggestions. > In light of the above comments, I will increase my score by one point. Just in case this somehow fell through the cracks, we believe the change (from 6 to 7) isn't reflected in your current score.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC