Decentralized Randomly Distributed Multi-agent Multi-armed Bandit with Heterogeneous Rewards

We study a decentralized multi-agent multi-armed bandit problem in which multiple clients are connected by time dependent random graphs provided by an environment. The reward distributions of each arm vary across clients and rewards are generated independently over time by an environment based on distributions that include both sub-exponential and sub-gaussian distributions. Each client pulls an arm and communicates with neighbors based on the graph provided by the environment. The goal is to minimize the overall regret of the entire system through collaborations. To this end, we introduce a novel algorithmic framework, which first provides robust simulation methods for generating random graphs using rapidly mixing Markov chains or the random graph model, and then combines an averaging-based consensus approach with a newly proposed weighting technique and the upper confidence bound to deliver a UCB-type solution. Our algorithms account for the randomness in the graphs, removing the conventional doubly stochasticity assumption, and only require the knowledge of the number of clients at initialization. We derive optimal instance-dependent regret upper bounds of order $\log{T}$ in both sub-gaussian and sub-exponential environments, and a nearly optimal mean-gap independent regret upper bound of order $\sqrt{T}\log T$ up to a $\log T$ factor. Importantly, our regret bounds hold with high probability and capture graph randomness, whereas prior works consider expected regret under assumptions and require more stringent reward distributions.

Paper

References (27)

Scroll for more · 15 remaining

Similar papers

Peer review

Reviewer H8V96/10 · confidence 2/52023-07-05

Summary

This paper studies the problem of multi-agent multi-armed bandits where the arm rewards generated for each agent is heterogeneous across agents, and the goal is to find the global optimal arm, where the reward of the global optimal arm is the mean of the mean reward observed by agents. In addition, this work considers that the agents are connected through an incomplete and time-varying topology. Both heterogeneous agent-specific reward and time-varying underlying topology make sense from a practical perspective and also add intellectually interesting technical challenges to the multi-agent MAB problem. The authors propose a series of algorithms for graph generation and bandit learning and then analyze the regret of the proposed algorithms under multiple different assumptions and in both problem-dependent and problem-independent settings. The regret bounds are interesting and (near) optimal in most cases.

Strengths

++ This paper addresses a very challenging setting of multi-agent MAB problem where reward generation is heterogeneous and the graph is time-varying. ++ The presentation of the paper is clear and solid and it clearly explains the challenges and missing points in the existing literature. ++ The proposed algorithms make sense intuitively. ++ The statements of the results sound solid (even though, I cannot confirm the details of the proof) and significant, with sufficient explanations in comparison with the literature.

Weaknesses

-- Overall, this paper is a solid work with novel and non-trivial contributions. Unfortunately, the technical part of this paper is beyond my core expertise, and I can't note on the correctness of the proofs. -- In the literature of multi-agent bandits, there is a substantial thread of the work that considers communication cost as a metric to improve upon. This work, however, does not touch at all the communication complexity in their study. Can authors comment on whether or not it makes sense to add communication cost analysis to their work? -- In this work, the authors, by default, assume that the goal of the game is to find the global optimal arm. However, one may suggest that in a multi-agent setting, each agent may care more about their local optimal arm. Can the authors comment on how the algorithms may or may not work when the objective is to find local optimal arms?

Questions

see two questions above.

Rating

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

n/a

Reviewer zM1d7/10 · confidence 4/52023-07-06

Summary

Update: I've read the authors' rebuttal as well as all the reviews from other reviewers. The authors' clarification are helpful and mitigate my concerns. I've raised the score from 6 to 7 and I think this manuscript is acceptable for publication. This paper aims at making contributions to decentralized multi-agent MAB problem research by studying both heterogeneous rewards (with heavy-tailed reward distributions) and time-varying random graphs, where the distributions of rewards and graphs are independent of time. It is claimed that this is the first work to consider this problem and to investigate it with heavy-tailed reward distributions. To be more specific, the paper investigates 1) heterogeneous sub-exponential and sub-gaussian distributed rewards and 2) random graphs including the possibly unconnected E-R model and random connected graphs, and applies them to the decentralized multi-agent MAB framework using a proposed UCB-type solution. It is considered as an extension of relevant existing literature (especially [Dubey et al.,2020] and and [Zhu et al., 2021b]) on decentralized multi-agent MAB research area.

Strengths

1. Paper presentation is clean and clear, and overall speaking, the manuscript is reader friendly and straightforward to follow. 2. There are quite nontrivial novel theoretical analysis results included in the paper, which seems to be solid. 3. There are quite a couple of helpful remarks/explanations comparing with existing literature.

Weaknesses

1. I'd like to see more more explanations/elaborations about the motivation/practical backgrounds why those types graphs are considered and what are some example real world applications that might have those types of graphs. For example, this paper consider the uniformly distributed connected graph generated using Algorithm 1, but could the author provide any example real world applications that have that kind of graphs? Especially about the part of adding extra minimum number of edges to make the randomly sampled graph to be connected . That seems to be a purely artificial/simulation based approach, and I don't obviously see much real world applications that demonstrate such behavior. The paper introduction briefly mention random failure graph in real world applications as motivation, but obviously random failed graphs have essential difference compared with such formulated graphs. More elaborations/explanations about the motivations/applications would be helpful to convince that such an artificial formulation is not just for convenient mathematical analysis, but also very beneficial to real world applications, rather than just a mathematical game. 2. In algorithm 1, I saw a constant probability "1/2" in the graph sampling part. Why this specific "1/2" is used here? Would any constant probability 0<p<1 work the same here for the analysis and theoretical results (as well as obtained bounds when comparing with literature)? It would be great if the authors could help clarify/confirm this in the manuscript. 3. I saw undirected graph is considered/assumed in this paper. How hard/difficult to consider the more generic scenario of directed graphs and what are the main challenges? It would be valuable to see some discussions on this for future research directions. 4. I'd like to see some scaling and complexity discussions. Is the proposed method naturally scalable to large multi-agent systems? It would be good to see some experiments and analysis on this, ideally comparing with relevant literature.

Questions

Please see my questions in my above comment when taking about the weakness.

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.

Area Chair abKL2023-08-18

Dear Reviewer, Please reply to the rebuttal and indicate whether it clears your concerns, or at least acknowledge whether you have read the response. This is important to the authors. Thanks, Your Area Chair

Reviewer zM1d2023-08-18

I've read the authors' rebuttal as well as all the reviews from other reviewers. The authors' clarification are helpful and mitigate my concerns. I've raised the score from 6 to 7 and I think this manuscript is acceptable for publication.

Reviewer 3jx97/10 · confidence 4/52023-07-07

Summary

This paper presents a novel algorithm for solving decentralized multi-agent multi-armed bandit problems for two graph dynamics under both sub-exponential and sub-gaussian reward distributions. Their whole algorithm framework can be divided into a Burn-in period and a Learning period, ensuring a process warm-up phase before getting UCB-type estimators. The paper includes detailed theorems and proofs in both the main manuscript and appendix with regret analysis.

Strengths

1. Authors fill in the gap in the multi-agent multi-armed bandit problem with the intersection between heavy-tailed reward distributions and heterogeneous settings with time-varying graphs. 2. The paper is clearly written and easy to follow.

Weaknesses

1. The paper is full of details but without a conclusion to re-emphasize the main statement and contribution of this paper. 2. Please refer to the questions

Questions

1. In line 368, is there any typo mentioning the same conditions as in Theorem 2 and $\textbf{4}$? 2. In line 312, the 3rd and 4th terms are exactly the same. Is there any typo? 3. Could you please elaborate more about how you get L1 in theorem 4?

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

4 excellent

Limitations

No

Reviewer NP2Q7/10 · confidence 3/52023-07-10

Summary

The paper studies stochastic multi-armed bandit over a network of agents communicating locally. Specifically, for each of the K arms, each agent has an associated underlying arm-mean. The agents can only communicate with their neighbors in each round to exchange information. The goal of the agents is to pull arms so that the regret with respect to the average arm-mean averaged over the agents is minimized. This is a challenging problem since the agents can only infer their arm mean from the observed rewards and must rely on information sharing to learn the average of the arm means averaged over the network. In order to model information sharing, the paper assumes that the communication graph over agents is a stochastic process that possibly changes over time. This allows for the information about an arm collected by an agent to percolate over the network so that all agents can potentially learn of the mean of an arm averaged over agents. However, the agents never have full information on the graph connectivity and at each time, can only observe who their neighbors are. In this scenario, the paper proposes an algorithm based on initial coordinated exploration and information sharing followed by UCB based on estimated network sample average over arms. The proposed algorithm is elegant and practical.

Strengths

A clear articulation of the problem and an elegant algorithm to solve it. Good analysis of the algorithm so that a reader can follow along. High probability bounds that do not average over the randomness in the graph generation process that is typical of results in this area.

Weaknesses

Lack of experimental evidence. Given one of the core contributions of the paper is an elegant algorithm, presenting synthetic simulations will substantiate the contributions of the paper.

Questions

Can the authors substantiate the insights using synthetic simulations? For example, it will be nice if the authors can - "visualize" the flow of information in the algorithm, - the impact of heterogeneity to the resulting regret - Impact of the graph generation process on regret. For example complete graphs is the easiest while a ring graph is the hardest? Perhaps the graph spectrum plays a role in the regret? Of-course these are a lot of changes, but will appreciate intuition/insights on these questions and will make the paper more complete. ***EDIT Post Rebuttal:*** -- I thank the authors for the thorough response and would encourage them to add it to the camera ready. I would further propose to add two related references to the additional list of references. Sankararaman et.al. Social Learning in Multi Armed Bandits, SIGMETRICS 2019 Chawla et.al. The Gossiping Insert-Eliminate Algorithm for multi-agent Bandits, AISTATS 2020

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

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

NA

Reviewer a3u47/10 · confidence 2/52023-07-24

Summary

This paper studies a multiplayer graph bandit setting. In particular, they consider heterogenous rewards in which the mean rewards for each agent are different, and are allowed to be both sub-gaussian and sub-exponential. They also consider random graphs between agents, which dictates which agents are allowed to communicate their rewards with each other. This graph structure is allowed to vary with time. They consider two types of graph generations as well as two types fo reward families for a total of 4 settings. They propose an algorithm with the following components (1) graph generation, (2) DrFed-UCB (burn in period), and (3) DrFed-UCB (learning period). Their regret analysis achieves optimal instance dependent regret on the order of log T with high probability.

Strengths

This paper considers a setting that is very general and interesting as well. For instance rewards can vary as well as the communication graphs which are generally assumed to be fixed. Furthermore, they allow for heavy tailed reward dsitributions where as standard bandit literature only considers subgaussian rewards. Their results seem to be very strong for this rather general setting.

Weaknesses

I think the organization of the settings as well as the key results can be better (Maybe each setting gets its own theorem latex environment). Moreover, I'd use section 3.2 to list the regret bounds, and outline the analysis in a section following that (theorem 1 can go in that section). Kind of related to this point I think the statements of theorem 2 , theorem 3, theorem 4 are too detailed. Maybe just state the regret bound and define the constants in paragraph below it or in the proof.

Questions

I'm not sure I understand S_1, S_2, and S_3. This is described in the first paragraph of sction 3.2 and it's worded quite confusingly. for isntance I'm not sure what M is referring to in describing Seettings 1.1, 1.2, 2.1, and 2.2. Also ,whcih setting is theorem 4 for? S_1, S_2 or S_3?

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

4 excellent

Presentation

3 good

Contribution

4 excellent

Limitations

The graph generation is limited to two distributions (ER and uniform). Furthermore, their algorithm can be costly in communication.

Authorsrebuttal2023-08-10

References: [1] A. Dubey and A. Pentland. Cooperative multi-agent bandits with heavy tails. In International Conference on Machine Learning, 2730–2739, 2020. [2] A. Frieze and M. Karo ́nski. Introduction to random graphs. Cambridge University Press, 2016. [3] M. Jerrum and A. Sinclair. The markov chain monte carlo method: an approach to approximate counting and integration. Approximation algorithms for NP-hard problems, pages 482–520,1996. [4] C. Gray, L. Mitchell, and M. Roughan. Generating connected random graphs. Journal of Complex Networks, 7(6):896–912, 2019. [5] M. Roughan, J. Tuke, and E. Parsonage. Estimating the parameters of the waxman random graph. In Algorithms and Models for the Web Graph: 16th International Workshop, WAW 2019, Brisbane, QLD, Australia, July 6–7, 2019, Proceedings 16, pages 71–86. Springer, 2019. [6] C. Wu, F. Wu, L. Lyu, Y. Huang, and X. Xie. Communication-efficient federated learning via knowledge distillation. Nature communications, 13(1):2032, 2022. [7] T. Li and L. Song. Privacy-preserving communication-efficient federated multi-armed bandits. IEEE Journal on Selected Areas in Communications, 40(3):773–787, 2022.

Reviewer a3u42023-08-10

Response to Authors

Thank you for taking the time to respond to my suggestions. I'm happy with the responses, references and the changes you plan on making.

Authorsrebuttal2023-08-17

Thank you for reviewing the responses and providing valuable feedback and suggestions. We will integrate the valuable suggestions and make the corresponding changes in the revised version of the paper.

Authorsrebuttal2023-08-10

Perspectives on numerical experiment details

Numerical Experiment Details Dataset generation: The process of data generation involves both reward generation and graph generation. First we generate different numbers of arms and clients, denoted as $K$ and $M$, respectively. Specifically, we generate rewards using both sub-gaussian and sub-exponential distributions, varying the mean values $\mu_i^m$ by random sampling within $[0,1]$ and introducing multiple levels of heterogeneity denoted as $h = \max_{i,j,m}|\mu_i^m-\mu_j|$. Additionally, we include multiple variance (proxy) $\sigma^2$ of the distributions by again random sampling. In terms of graph generation, we generate E-R models with varying values of $c$, and we generate random connected graphs using Algorithm 1. Furthermore, as elaborated in A2, we would consider the generation of random ring graphs or star graphs. Models and benchmarks: We will implement the proposed algorithms. We consider \citep{dubey2020cooperative} as the benchmark, since the work assumes heavy-tailed rewards and deterministic graphs in the analysis, and additionally it conducts experiments with a fixed E-R model. Meanwhile, we also compare the new method with~\citep{zhu2021federated} which focuses on deterministic graphs and sub-gaussian rewards and motivates our work. Model evaluation: The evaluation metric is the regret measure as defined in the paper. An additional possibility is to include communication cost as another performance measure, based on the reviewers' suggestions and existing literature. This will help us understand the trade-off between regret and communication cost under different models.

Reviewer NP2Q2023-08-11

Thank you for the response

I thank the authors for the thorough response and would encourage them to add it to the camera ready. I would further propose to add two related references to the additional list of references. Sankararaman et.al. Social Learning in Multi Armed Bandits, SIGMETRICS 2019 Chawla et.al. The Gossiping Insert-Eliminate Algorithm for multi-agent Bandits, AISTATS 2020

Authorsrebuttal2023-08-17

Thank you for reviewing the responses and for sharing the valuable suggestions and references. We will incorporate the discussions and references [5, 6] into the paper and benchmarks, as elaborated in the following paragraphs verbatim. The works [5, 6], in addition to establishing the elegant algorithms and optimal regret bounds that explicitly characterize the dependency of regret bounds on the conductance of the graph, also pave the way for integrating communication efficiency and privacy guarantee optimization into our settings with random graphs and heterogeneous rewards. The characterization of the trade-off between regret and communication cost in the regret bound theoretically demonstrates the improvement resulting from replacing ring graphs with complete graphs. The techniques include asynchronous communication protocols and communication cost optimization, in line with both federated learning (the number of transmitted estimators is $O(M)$ or $\log (MK)$, instead of the common $O(MK)$) and decentralized multi-agent MAB (the number of communication rounds having a budget of $O(\log T)$ or $o(T)$, instead of $O(T)$). We appreciate the references and have added them to the communication efficiency discussion that is also suggested by reviewer a3u4 and reviewer H8V9. In the meantime, transmitting arm selections, rather than reward estimators based on historical arm pulls, significantly enhances client privacy. Modifications to Introduction Line 47: In the field of decentralized multi-agent MAB, it is commonly assumed that the mean reward value of an arm for different clients is the same, or equivalently, homogeneous. This assumption can be seen in works such as [..., 3, 5, 6]. Line 66: Besides rewards, the underlying graph assumptions are essential to the decentralized multi-agent MAB problem, as increased communication among clients leads to better identification of global optimal arms and smaller regret. The existing works [5, 6] relate regret with graph structures and characterize the dependency of regret on the graph complexity with respect to conductance. When considering two special cases, [6] demonstrates the theoretical improvement achieved by circular ring graphs compared to complete graphs, and [3] numerically shows that the circular ring graph presents the most challenging scenario with the largest regret, while the complete graph is the simplest. Modifications to Remark (Specification of the parameters) Line 335: We observe that the regret bound is dependent on the transition kernel $\pi$ and the spectral gap $\lambda$ of the underlying Markov chain associated with $\pi$. This indicates the significance of graph complexities and distributions within the framework of the random graph model when deriving the regret bounds, in a similar manner as the role of graph conductance in the regret bounds established in [5, 6] for time-invariant graphs. Modifications to our experimental details Models and benchmarks: We will implement the proposed algorithms. In term of heavy-tailed rewards, we will consider [1] as the benchmark, since the work assumes heavy-tailed rewards and deterministic graphs in the analysis, and additionally it conducts experiments with a fixed E-R model. Meanwhile, we will also compare the new method with [2] which focuses on deterministic graphs and sub-gaussian rewards and motivates our work. We will also include [5, 6] as benchmarks, where the regret bounds in [5, 6] fully characterize the dependencies on the sub-optimality gap and the graph complexity in terms of conductance. These works and [3] also consider communication efficiency and privacy, and the experiments will also cover complete graph scenarios. Variables in the settings: We will explore various settings to analyze the tightness of the regret bounds' dependencies on a range of factors. These factors include the degree of setting heterogeneity ($h$), the distributions of graphs with different complexities ($f$), the number of clients ($M$), the number of arms ($K$), and the reward distributions ($\sigma^2$, $\mu$, and $s$), where $s$ denotes whether the distribution is light-tailed or heavy-tailed. Evaluation: The evaluation metric will be the regret measure as defined in the paper. We will also include the communication cost as another performance measure, which is thoroughly considered in [3, 5, 6], based on the reviewers' suggestions and existing literature. This will help us understand the trade-off between regret and communication cost under different models. Meanwhile, the privacy guarantee will also be regarded as an evaluation metric, as optimized in [5, 6], following the privacy measure defined in [3] or a newly defined measure. Additionally, runtime can provide insights into the time complexity of the models. [5] Sankararaman et.al. Social Learning in Multi Armed Bandits, SIGMETRICS 2019 [6] Chawla et.al. The Gossiping Insert-Eliminate Algorithm for multi-agent Bandits, AISTATS 2020

Authorsrebuttal2023-08-10

Perspectives on the concluding section and the alternative of L

Concluding section: In this paper, we consider a decentralized multi-agent multi-armed bandit problem in a fully stochastic environment that generates time-varying random graphs and heterogeneous rewards following sub-gaussian and sub-exponential distributions, which has not yet been studied in existing works. To the best of our knowledge, this is the first theoretical work on random graphs including the E-R model and random connected graphs, and the first work on heterogeneous rewards with heavy tailed rewards. To tackle this problem, we develop a series of new algorithms, which first simulate graphs of interest, then run a warm-up phase to handle graph dynamics and initialization, and proceed to the learning period using a combination of upper confidence bounds (UCB) with a consensus mechanism that relies on newly proposed weight matrices and updates, and using a stopping time to handle randomly delayed feedback. Our technical novelties in the results and the analyses are as follows. We prove high probability instance-dependent regret bounds at the order of $\log T$ in both sub-gaussian and sub-exponential cases, consistent with the regret bound in the existing works that only consider the expected regret. Moreover, we establish a nearly tight instance-free regret bound of order $\sqrt{T}\log T$ for both sub-exponential and sub-gaussian distributions, up to a $\log T$ factor. We leverage probabilistic graphical methods on random graphs and draw on theories related to rapidly mixing Markov chains, which allows us to eliminate the doubly stochasticity assumption through new weight matrices and a stopping time. We construct new network-wide estimators and invent new concentration inequalities for them, and subsequently incorporate the seminal UCB algorithm into this distributed setting. Alternative of $L$: Alternatively, we may rephrase the expression for $L$ as follows - by keeping the repeated term ($\frac{4K\log_{2}T}{c_0}$ which is another repetitive term that was omitted in the earlier expression; it is associated with the edge probability $c$, which defines the gap between consecutive communications and remains consistent across both $s_1$ and $s_2$). $$ L = \max{L_{s_1},L_{s_2},L_{s_3}} $$ where $L_{s_1} = \max(\frac{{}\ln{\frac{2T}{\epsilon}}}{2\delta^2}, \frac{4K\log_{2}T}{c_0})$, $L_{s_2} = \max(\frac{\ln{\frac{\delta}{10}}}{\ln{p^*}} +25\frac{1+\lambda}{1-\lambda}\frac{\ln{\frac{2T}{\epsilon}}}{2\delta^2}, \frac{4K\log_{2}T}{c_0})$, and $L_{s_3} = \max(\frac{\ln{\frac{\delta}{10}}}{\ln{p^*}} +25\frac{1+\lambda}{1-\lambda}\frac{\ln{\frac{2T}{\epsilon}}}{2\delta^2}, \frac{K\ln(\frac{MT}{\epsilon})}{c_0\ln(\frac{1}{1-\frac{2\log M}{M-1}})})$ are the length of burn-in period of setting $s_1, s_2$ and $s_3$, respectively. There are two terms we have not introduced, $\frac{{}\ln{\frac{2T}{\epsilon}}}{2\delta^2}$ and $\frac{K\ln(\frac{MT}{\epsilon})}{c_0\ln(\frac{1}{1-\frac{2\log M}{M-1}})}$ where the first is related to the concentration inequality of the empirical matrix $P_t$ in the E-R model and the second term is related to the gap between two consecutive communications when $c$ is small in setting $s_3$ and $S_3$ where $M > 11$.

Reviewer 3jx92023-08-15

Thank you for your response and clarification. The conclusion section will definitely make the work more complete.

Authorsrebuttal2023-08-17

Thank you for reviewing the responses and providing valuable suggestions and comments. We will incorporate the concluding section verbatim into the revised paper.

Authorsrebuttal2023-08-10

Perspectives on directed graphs and complexity analyses

Q3. How hard/difficult to consider the more generic scenario of directed graphs and what are the main challenges? This paper aligns with most existing literature on random graph generation and on decentralized distributed multi-agent MAB. A recent separate line of research focuses on strongly connected graphs. We highlight differences and challenges for directed graphs across key aspects. Graph Generation: In the E-R model, individual directed edge sampling occurs via a Bernoulli distribution with $c$ probability. The sampling frequency adjusts to $M(M-1)$ from $\frac{M(M-1)}{2}$ in undirected graphs. For random connected graphs, Algorithm 1 requires adjustment to account for a uniform distribution over all strongly connected graphs. Edge selection involves sampling all arcs, followed by verification to ensure the resulting graph is strongly connected. This change preserves prior time complexity. The convergence rate analysis in this strongly connected variant would require additional effort, though the techniques employed should be similar to the current analyses applied to undirected graphs, following Propositions 3 and 4 in Appendix. We can demonstrate that the acceptance ratio $\alpha_t$ of the underlying Metropolis-Hastings sampling remains unchanged from the current version. A challenge lies in establishing a new transition kernel $\pi(\cdot,\cdot)$ for strongly connected graphs, as in Proposition 4. This points to future work, leveraging existing analytical techniques in this work. Main Algorithm: Asymmetry in empirical matrices $P_t$ and $P^{\prime}_t$ necessitates modifications. Updates for client $m$ switch to ${P_t(j,m)}$ and ${P^{\prime}_t(j,m)}$ from ${P_t(m,j)}$ and ${P^{\prime}_t(m,j)}$. Network-wide estimators $\Tilde{\mu}$ include arcs, involving arc-related substitutions. Integrating asymmetrical information into the algorithm is a challenge that can enhance consensus for strongly connected graphs. Regret Analysis: Taking Theorem 2 as an example, achieving a regret bound of $O(\log T)$ requires: 1) empirical matrix convergence, 2) client consensus: consensus on the number of pulled arms, and 3) concentration inequalities for network-wide estimators. Step 2 involves defining sufficient conditions on strong connectivity in the context of the E-R model. The remaining steps align with the current analysis framework. Q4. I'd like to see some scaling and complexity discussions. Is the proposed method naturally scalable to large multi-agent systems? It would be good to see some experiments and analysis on this, ideally comparing with relevant literature. A4. The method is theoretically suitable for multi-agent systems of any size due to universal regret bounds regardless of $M$. For Erdős-Rényi (E-R) graphs, time complexity is $O(M^2)$, but independent edge sampling allows parallelization for larger systems, notably accelerating the process. For random connected graphs, initialization complexity is $O(M^3)$ (discussed in A1), but its impact on overall efficiency is limited given the time horizon's context. At each step, graph generation's time complexity is $O(M^2 + M + |E_t|) \leq O(M^2)$ where the first term accounts for edge selection, and the second and third terms are for graph connectivity verification using Algorithm 1 (benefitting from the use of Markov chains for graph generation). The main algorithm's stages result in $O(M^2 + MK)$ overall complexity, where the $MK$ complexity results from the stages like UCB (Upper Confidence Bound) computations and estimator updates, while $M^2$ arises from transmissions and graph generation. [Zhu et al., 2021b] proposes $Gossip_{UCB}$ method for decentralized multi-agent MAB with time-invariant connected graphs which exhibits a time complexity of $O(MK + M^2)$. Our method is consistent with this work, and beyond that, we account for more complicated random graphs, which surprisingly does not bring an additional time complexity ($M^3$ in the initialization step in the worst-case scenario to obtain a connected graph is also applicable to this work). The $Decentralized , MP-UCB$ algorithm designed for time-invariant connected graphs and homogeneous heavy-tailed rewards in [Dubey et al.,2020] has $O(MK + M^2)$ complexity per step, with initialization at $O(3^{\frac{M}{3}})$. Ours retains complexity order, with significantly lower initialization and extension to random graphs. Most procedures operate synchronously, except random connected graph generation ($O(M^2)$). Our approach balances parallelism and random connected graph generation, yielding reasonable $O(M^2)$ complexity for exponentially large sample spaces. Space complexities align with existing methods at $O(M^2 +MK)$ due to estimator storage. Regarding experimentation, while we were not able to complete the experiments within the response period, we plan numerical studies to assess runtime and performance against existing methods.

Reviewer H8V92023-08-14

Response

Thanks for your response. It would be great if you could add that communication analysis to the updated version. Also, there are some works in the context of multi-agent bandits with homogeneous rewards that achieve much lower communication costs (see below), where the algorithm design follows some other approaches, e.g., leader-follower, so, considering those might lead to better trade-offs between regret and communication cost. [1] Wang, Po-An, et al. "Optimal algorithms for multiplayer multi-armed bandits." International Conference on Artificial Intelligence and Statistics. PMLR, 2020. [2] Wang, Xuchuang, et al. "Achieving Near-Optimal Individual Regret & Low Communications in Multi-Agent Bandits." The Eleventh International Conference on Learning Representations. 2022.

Authorsrebuttal2023-08-17

Thank you for reviewing the response and sharing the valuable advice and references. We will incorporate the communication cost analysis and the new references into the revised version of the paper, as detailed in the following paragraphs. Modifications to the introduction in the paper Line 47: In the field of decentralized multi-agent MAB, it is commonly assumed that the mean reward value of an arm for different clients is the same, or equivalently, homogeneous. This assumption can be seen in works such as [..., 1, 2]. Modifications to Section 3 (including [1, 2]; the references [5, 6] suggested by reviewer NP2Q) Communication cost analysis: Assuming a constant cost of establishing a communication link, as defined in [1, 3], denoted as $c_1$, the communication cost $C_T$ can be calculated as $C_T = c_1 \cdot \sum_{t=1}^T |E_t|$, which is proportional to $\sum_{t=1}^T |E_t|$. Alternatively, following the framework proposed in [2, 4, 5, 6], the communication cost can be defined as the total number of communications among clients, which can be represented as $C_T = \sum_{t=1}^T |E_t|$, similar to the previous definition. Regarding the quantity $C_T = \sum_{t=1}^T |E_t|$, the number of edges $E_t$ could be $O(M)$ for sparse graphs, and at most $O(M^2)$. In the random graph model, the expected number of edges is $\frac{M(M-1)}{2}c$, which implies $O(M^2)$ in the worst case scenario and that the total communication cost, in a worst-case scenario, is of order $O(TM^2)$. This analysis holds also for the random connected graph case where $c$ represents the probability of having an edge. This cost aligns with the existing work of research on decentralized distributed multi-agent MAB problems without a focus on communication cost, where edge-wise communication is a standard practice. Added text to Conclusions and Future Work Recent advancements have been made in reducing communication costs with respect to the dependency in multi-agent MAB with homogeneous rewards (in the generalized linear bandit setting [4], the groundtruth of the unknown parameter is the same for all clients), such as achieving $O(\sqrt{T}M^2)$ in [4] for centralized settings or $O(M^3\log T)$ through Global Information Synchronization (GIS) communication protocols assuming time-invariant graphs in [3]. Likewise, [5, 6] improve the communication cost of order $\log T$ or $o(T)$ through asynchronous communication protocols and balancing the trade-off between regret and communication cost. More recently, [2] establishes a novel communication protocol, TCOM, which is of order $\log \log T$ by means of concentrating communication around sub-optimal arms and performing aggregation of estimators across time steps. Furthermore, [1] develops a new leader-follower communication protocol, which selects a leader that communicates to the followers. Here the communication cost is independent of $T$ which is much smaller. The incorporation of random graph structures and heterogeneous rewards introduces its own complexities, which poses challenges to reductions in communication costs. These great advancements introduce a promising direction for communication efficiency as a next step within the context herein. References: [1] Wang, Po-An, et al. "Optimal algorithms for multiplayer multi-armed bandits." International Conference on Artificial Intelligence and Statistics. PMLR, 2020. [2] Wang, Xuchuang, et al. "Achieving Near-Optimal Individual Regret \& Low Communications in Multi-Agent Bandits." The Eleventh International Conference on Learning Representations. 2022. [3] T. Li and L. Song. Privacy-preserving communication-efficient federated multi-armed bandits. IEEE Journal on Selected Areas in Communications, 40(3):773–787, 2022. [4] C. Li and H. Wang. Communication efficient federated learning for generalized linear bandits. Advances in Neural Information Processing Systems, 35:38411–38423, 2022. [5] Sankararaman et.al. Social Learning in Multi Armed Bandits, SIGMETRICS 2019 [6] Chawla et.al. The Gossiping Insert-Eliminate Algorithm for multi-agent Bandits, AISTATS 2020

Program Chairsdecision2023-09-21

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC