No-Regret Online Prediction with Strategic Experts

We study a generalization of the online binary prediction with expert advice framework where at each round, the learner is allowed to pick $m\geq 1$ experts from a pool of $K$ experts and the overall utility is a modular or submodular function of the chosen experts. We focus on the setting in which experts act strategically and aim to maximize their influence on the algorithm's predictions by potentially misreporting their beliefs about the events. Among others, this setting finds applications in forecasting competitions where the learner seeks not only to make predictions by aggregating different forecasters but also to rank them according to their relative performance. Our goal is to design algorithms that satisfy the following two requirements: 1) $\textit{Incentive-compatible}$: Incentivize the experts to report their beliefs truthfully, and 2) $\textit{No-regret}$: Achieve sublinear regret with respect to the true beliefs of the best fixed set of $m$ experts in hindsight. Prior works have studied this framework when $m=1$ and provided incentive-compatible no-regret algorithms for the problem. We first show that a simple reduction of our problem to the $m=1$ setting is neither efficient nor effective. Then, we provide algorithms that utilize the specific structure of the utility functions to achieve the two desired goals.

Paper

References (23)

Scroll for more · 11 remaining

Similar papers

Peer review

Reviewer oxKD6/10 · confidence 3/52023-07-02

Summary

This submission studies the problem of online decision-making using predictions given by experts with incentives to behave strategically. In particular, at each time-step, K experts hold a belief about a binary outcome. Each expert reports a prediction to a learner, who picks a set of m experts. The learner then suffers a loss which is a function of both the binary outcome and the chosen experts' true beliefs about the outcome. The goal of the learner is to achieve no-regret with respect to the best-in-hindsight choice of m experts while maintaining incentive compatibility, i.e. reporting their true belief as their prediction is a weakly-dominant strategy for each expert at every time-step. The authors study two settings: one in which the learner has a modular utility function and one in which their utility function is submodular. An existing algorithm (WSU, [6]) for the 1-expert version of this problem (i.e. the version of the problem where the learner selects one expert at each time-step) is capable of getting sublinear regret in both settings, albeit at the cost of exponential runtime. For modular utility functions, authors show how to use a variant of the well-known Follow the Perturbed Leader (FTPL) algorithm under a sufficient condition for the perturbation distribution to guarantee both no-regret and approximate incentive compatibility. For submodular utility functions, the authors use an "online distorted greedy algorithm" to obtain no-regret while maintaining incentive compatibility constraints exactly. At a high level, the online distorted greedy algorithm runs m algorithms for the strategic 1-expert problem concurrently. At each time-step, each of the m sub-algorithms selects an expert, and the learner uses these m experts as their selection. Based on the loss the learner receives, they set the loss of each sub-algorithm in a particular way. The authors use the WSU algorithm of [6] as their sub-algorithm. Along they way, they derive an adaptive regret bound for the WSU algorithm, which may be of independent interest. Finally, the authors empirically evaluate their two algorithms on a dataset from a FiveThrityEight forecasting competition to predict the outcomes of games in the 2022-2023 NFL football season. They find that both algorithms obtain similar performance in this setting.

Strengths

While others have studied the (non-strategic) m-expert problem, as well as the strategic 1-expert problem, the authors are the first to study the strategic m-expert problem. This setting is well-motivated by applications such as forecasting competitions. While the algorithms presented by authors are not particularly novel, their application to the strategic m-expert problem requires non-trivial theoretical analysis, as well as some new ideas, such as (1) the sufficient condition for the perturbation distribution to guarantee approximate incentive compatibility in the modular setting, and (2) the adaptive regret bound for the WSU algorithm used as a subroutine for the submodular setting. While I would normally consider the experimental results on real-world data as a strength, their impact is limited due to the lack of relevant comparisons (see Weaknesses for more details). Finally the writing is clear, which makes the authors' contributions easy to understand.

Weaknesses

The lack of any sort of lower bound is a weakness, especially since (as the authors point out) the loss function they consider is exp-concave, and thus the regret rates of Hedge in the non-strategic version of the problem where one expert is selected scale only logarithmically in the number of experts (and independently of the time horizon). While it may indeed be the case that going from constant-in-T to T^{1/2} regret rates is the price to pay for incentive compatibility in the m-expert problem, this submission presents no evidence either for or against this claim. It would be nice to have provide some background on the current state of the (non-strategic) m-expert problem in the related work. Finally, the experimental results would be strengthened considerably if (1) the theoretical rates and (2) the regret of the naive application of the WSU algorithm were also plotted. Currently, it is not easy to see how well the algorithms are actually doing without the relevant baselines. Additionally, it would be interesting to see how closely the empirical performance of the algorithms match their corresponding theoretical bounds.

Questions

What are the runtimes of both algorithms? What regret rates are obtained for the non-strategic m-experts problem with squared loss?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

The authors have adequately addressed the limitations of their work.

Reviewer Aodd6/10 · confidence 3/52023-07-05

Summary

This paper generalizes the problem of online binary prediction with expert advice, where at each round the learner can pick m>1 experts and the overall utility is a modular or submodular utility of the chosen experts. The experts are strategic and wish to be selected by the algorithm as often as possible hence they may misreport their beliefs about the events. They design algorithms that are incentive-compatible and achieve sublinear regret in the hindsight. Previous work has studied this problem for m=1. In the case of m>1, previous work have focused on designing no-regret algorithms and did not take into consideration the incentive-compatibility issues. Their algorithm builds on a prior work that studies the FTPL algorithm for the m-expert problem with modular utility functions and derive conditions for the perturbation function to guarantee approximate incentive compatibility. In particular, they show that FTPL while Gaussian perturbations is not incentive-compatible, Laplace or hyperbolic noise distribution guarantees approximate incentive-compatibility. Moreover, they propose an algorithm that builds upon online monotone submodular maximization s.t. matroid constraints that takes m incentive-compatible for standard experts problem (i.e. m=1) and outputs their combined prediction. Compared to the first algorithm, this one achieves exact incentive-compatibility however, at the price of an extra \sqrt{m} term in the regret bound.

Strengths

The paper is written nicely and is easy to follow. The problem seems to be well-motivated. I did not go through the proofs. The second algorithm is more straightforward, however, I think it is nice that they included it and compared it with their first algorithm.

Weaknesses

It might be helpful to add a discussion on the regret lower bound for this problem.

Questions

Minor comments: The citations need to be fixed, they do not show the authors' name, need to use \citet. Line 268, can you expand on why this holds? Question : What is the lower bound for the regret of this problem? Can you give a lower bound that is stronger than the case for m=1?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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.

Reviewer xXJd6/10 · confidence 3/52023-07-09

Summary

- The paper studies the problem of online prediction with expert advice. In their setting, a learner at each round selects $m$ experts, and the loss is determined by a submodular or modular function of the beliefs reported by the selected experts. Each expert, being strategic, may intentionally misrepresent their beliefs to maximize their chances of being selected in the next round. The goal of the learner is to design no-regret algorithms that incentive experts to report their beliefs truthfully (also called IC). - To tackle this problem, the authors first present an inefficient reduction to the case when $m=1$. Subsequently, they propose two efficient algorithms that are for the specific types of loss functions considered. - In the case of the modular loss function, the authors investigate Follow-The-Perturbed-Leader (FTPL) and propose a general condition for the perturbation distribution that ensures approximate IC. - For the submodular loss function,they propose a simpler algorithm named the online distorted greedy algorithm. This algorithm not only guarantees exact IC but also attains the optimal approximation ratio.

Strengths

- The paper studies an important and generic problem of online prediction with strategic experts that may find its applications in many scenarios. - The proposed algorithms in this paper are not only easy to implement and also enjoy good theoretical guarantees of being (approximate) IC and no regret.

Weaknesses

1. Several claims may have some issues, see eg Questions 3, 5, 6. 2. The presentation could be improved. For instance, Section 1.2 titled "Contributions" seems to contain some elements of background and motivations which might be more appropriately placed in the introduction of the paper. Moreover, it would be beneficial to include more discussions following important definitions (like Def 1) and theorems to provide more intuitions. 3. In the experimental section, the authors plot the regret curve that is not normalized. This makes it hard to interpret, as unnormalized regret curves don't explicitly demonstrate the rate at which the algorithms learn. Additionally, to reflect the difference between reported and true beliefs, a uniformly random value within the range guaranteed by Theorem 3 is added, but this theorem doesn't ensure a uniform distribution for this difference. Finally, incorporating a comparison of the proposed algorithms with and aggregated predictions of FiveThirtyEight could potentially enrich the analysis.

Questions

1. Regarding Def 1: - Is this a new concept being proposed, or has it previously been studied in previous works? - The experts are long-standing in this setting, but the focus of the paper is on myopic IC. Although the authors mentioned that the analysis of FTPL could be extended to the more general setting of maximizing a conic combination of probabilities of being chosen at all subsequent rounds, is it also true for the online distorted greedy algorithm? What's the main modification/challenge of extending to non-myopic experts? - Would achieving IC become impossible if the definition is for all $r_t$, rather than the expectation over Bern( $b_{i,t}$ )? 2. In the "forecasting competitions" example described in Section 3.1, it seems that the loss function is not captured by the modular function studied in this paper, given it is normalized by $|S_t|$ rather than $m$. 3. About Thm 1, could you clarify why the reduction preserves IC? In particular: - How is the belief of each meta-expert defined in this context? Is the new loss function $\ell_S$ still proper? How to define IC when the belief of each expert is multi-dimensional? - Even though WSU is IC for the $m=1$ problem, it only guarantees every single expert reports truthfully when fixing other's reports. In particular, it doesn't rule out the possibility that a group of experts form a coalition and misreport together to increase the sum of their probabilities of being selected, which is exactly what would happen when a single expert misreports in the $m$-expert problem -- it results in the simultaneous misreporting of $\binom{K}{m-1}$ meta-experts. 4. Thm 3 guarantees approximate IC in the sense of $|p^\star\_{i,t}-b\_{i,t}|$ being small. However, a more common notion of approximate IC is through the difference of utility, which, applies to this setting, is the difference in $\pi_{i,t+1}$ when $p^\star_{i,t}$ and $b_{i,t}$ are reported respectively. Could the authors include a discussion about this alternative notion? 5. Combining Thm 3 and Thm 5, it seems that as $B$ approaches 0, the FTPL algorithm becomes more IC and also yields a smaller regret. However, this contradicts the intuition that if the learner's decisions are entirely noise-driven (i.e., $B=0$), the regret should be high. Could the authors comment on this? 6. On lines 297-298, you mention that the $(1-\frac{c_f}{e})$ approximation factor is optimal. Is this claim supported by a lower bound? If so, could you point to the relevant references?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

3 good

Contribution

3 good

Limitations

Yes.

Reviewer 9kMc6/10 · confidence 3/52023-07-09

Summary

The paper broadly focuses on the design and analysis of no-regret and incentive-compatible algorithms for the m-experts problem (pick a subset $S_t \subseteq [K]$ of $m$ experts in each round $t$ to obtain a utility value $f_t(S_t)$ which is a function of the losses $\ell_{i,t} \in [0,1]$ for all $i \in S_t$), with specific utility functions (on the space of m-sized subsets of experts) in each round which are either modular or sub-modular. * They try to improve on the inefficient baseline of using the no-regret, incentive-compatible WSU algorithm (Freeman et all 2020) on the set of $\binom{K}{m}$ meta-experts corresponding to sets of $m$ experts. * They analyze the FTPL algorithm in the modular utility function case where the perturbation distribution is zero-mean symmetric from an exponential family. They propose a sufficient-condition for this distribution (related to bounded hazard rate) which implies that FTPL would be no-regret and approximately incentive-compatible. In particular, they show that this condition is satisfied for the Laplace distribution and the symmetric hyperbolic distribution etc. * They propose and analyze an "online distorted greedy algorithm" (based on earlier work by (Harvey et al. 2020) on online submodular maximizination) for the submodular utility function case. They show that the algorithm has sublinear $\alpha$-regret bounds and is $\alpha$-approximately-incentive compatible. where $\alpha$ may be $ < 1$ for general submodular functions (depending on the "curvature") while $\alpha = 1$ in the modular case. This of course implies that the algorithm has no-regret in the usual sense and is fully incentive compatible in the modular case, but at the expense of more computation and a slightly worse regret bound compared to FTPL.

Strengths

* The online $m$-experts problem is well-motivated in the introduction, as is the importance of incentive compatibility for the problem. * The algorithms proposed (FTPL and online distorted greedy) are natural for the problem and build nicely on previous work. All the appropriate details are given, along with the code and experiments. * The proposed sufficiency condition for the perturbation distribution of FTPL is fairly general, and may be of independent interest in the study of incentive compatibility. * Sufficient proof-sketches are given for quite a few of the results in the main paper itself, and more detail is provided in the supplementary material.

Weaknesses

* The FTPL analysis applies to a single modular utility function defined in the paper (Line 136). Of course, this utility function is fairly natural for the problem. * The necessity of "Condition 1" for getting approximate-incentive-compatibility with FTPL is not explored. * The experimental evaluation provided is fairly limited (of course, the thrust of the work is theoretical).

Questions

None

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, 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

4 excellent

Presentation

3 good

Contribution

3 good

Limitations

Not applicable.

Reviewer oxKD2023-08-11

Thanks for your detailed reply. Please allow me to clarify my question about the runtime. In particular, I am asking about the theoretical runtime in Big O notation, as opposed to wall clock time.

Authorsrebuttal2023-08-11

Running Time of the Proposed Algorithms

Thanks for the clarification. At each round, the FTPL algorithm simply requires picking the $m$ experts (among the $K$ available experts) with the smallest noisy losses. We implemented this using a binary heap whose running time is $\mathcal{O}(K+m\ln K)$ per round. The online distorted greedy algorithm uses $m$ instances of the WSU algorithm for the $1$-expert problem and simply outputs the union of the experts chosen by these algorithms. At round $t\in[T]$, The WSU algorithm takes $\mathcal{O}(K)$ to compute the probabilities $\pi_{i,t}$ and pick one expert according to these probabilities. Therefore, the running time of the online distorted greedy algorithm is $\mathcal{O}(mK)$ per round. We are happy to answer any further questions.

Reviewer oxKD2023-08-11

Shouldn't the runtime of the online distorted greedy algorithm be random, due to the fact that it may be necessary to re-sample experts (as stated in Line 294)? If so, is $\mathcal{O}(mK)$ the *expected* per-round runtime?

Authorsrebuttal2023-08-11

First, note that we did not require the algorithm to do resampling to obtain the theoretical bound. Therefore, the proven regret bound holds even if there are repeated experts that are chosen by the $m$ instances of the WSU algorithm. In this case, we can simply set the loss of the repeated experts to be 1. That being said, assume that at round $t\in[T]$, the WSU instance $\mathcal{A}\_j$ should pick an expert and the set of experts $\\{v\_{1,t},\dots,v\_{j-1,t}\\}$ are already chosen by $\mathcal{A}\_1,\dots,\mathcal{A}\_{j-1}$. In this case, the number of samples $\mathcal{A}_j$ needs to take to obtain a new expert follows a geometric distribution whose expected value is $\frac{1}{p}$ where $p$ is the probability of success (i.e., probability of picking one of the $K-j+1$ experts that are not already chosen). However, in practice, there are ways to simulate this step without actually performing repeated sampling. This question was addressed in Section 4.6.1 (title: "Avoiding duplicate actions") of [this paper](http://reports-archive.adm.cs.cmu.edu/anon/anon/home/ftp/usr0/ftp/2007/CMU-CS-07-171.pdf) who first introduced the online greedy algorithm for maximizing submodular functions.

Reviewer oxKD2023-08-11

Thanks for the reply. I encourage the authors to include such a discussion on runtime (as well as the related work on the non-strategic $m$-expert problem) in the revision. With that being said, most of my concerns have been addressed by the authors' replies, and so I will raise my score from 5 to 6.

Authorsrebuttal2023-08-11

Thanks for taking the time to read our rebuttal, we are glad that our response has addressed your concerns. We will make sure to use the additional content page in the final version of the paper to include the discussion on runtime and also the related work on the non-strategic $m$-experts problem.

Reviewer xXJd2023-08-15

Thank you very much for the detailed response. Given that most of my concerns have been addressed, I'm happy to raise my score from 5 to 6.

Authorsrebuttal2023-08-15

Thanks for taking the time to go over our response, we are glad that we were able to address most of your concerns.

Reviewer Aodd2023-08-16

Responding to rebuttal

Thank you for your detailed response. I understand that addressing the lower bound is not a necessity for this paper and it can be studied in future work. It might be helpful to add your conjecture for the lower bound and raise it as an open question.

Authorsrebuttal2023-08-16

Thanks for taking the time to read our response, we will make sure to use the additional content page in the final version of the paper to add the discussion about lower bounds.

Reviewer 9kMc2023-08-21

Response

Thank you for the clarifications in this and in the global rebuttal. I do believe incorporating some of those points into the final version of the paper will certainly make it better. I will keep the score for now but will take the clarifications in consideration if further discussion with the AC becomes necessary.

Authorsrebuttal2023-08-21

Thanks for taking the time to read our response, we will make sure to use the additional content page in the final version of the paper and incorporate the points you mentioned in your review (particularly the necessity of Condition 1).

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC