Population game dynamics describe how aggregate behavior evolves in response to payoff signals. In many applications, the payoff function may change, and the goal is to predict the population trajectory induced by a new payoff design before trajectory data under that design are available. We study this problem as payoff-aware prediction of population game dynamics. Standard dynamics learning methods fit the motion observed under training payoffs, but do not separate the payoff-dependent incentive from the payoff-independent response structure needed for prediction under new payoffs. Building on the existing framework of Riemannian game dynamics, we propose PA-RmD, a payoff-aware learning method that keeps payoff information explicit and learns the payoff-independent response structure from data. Theoretically, we show that the learned model preserves fixed-payoff Nash equilibria and derive finite-time one-step prediction bounds under shared-response and payoff-coverage conditions. Empirically, we evaluate PA-RmD on synthetic zero-sum and potential games, as well as a YouTube Trending multi-country dataset, showing that our method performs well under changing payoff designs.
Introduction
Population games study systems with many agents who repeatedly choose among a finite set of strategies. In such games, a payoff function specifies the benefit agents receive from each strategy, and the associated population game dynamics describe how the aggregate behavior of the population changes under these payoffs. This viewpoint is standard in large-population models of congestion systems, markets, and other settings in which individual identities are less important than aggregate behavior [38, 15, 35]. The same aggregate perspective has also been adopted in models of online content platforms [49, 23], where content creators repeatedly adjust what content to produce in response to platform-provided payoff signals, such as exposure or monetary rewards [50, 9, 47, 10].
Predicting such population game dynamics can provide important guidance for decision-making and mechanism design. In many design settings, the payoff function is not fixed: a planner or platform may alter the incentives faced by the population, and these changes can substantially reshape the induced population dynamics. Prediction in this setting must therefore be payoff-aware: the model should take a candidate payoff function as input and forecast the population trajectory it would induce. We refer to this problem as payoff-aware prediction of population game dynamics. The challenge is that the prediction is counterfactual. For example, a content platform may want to know whether a new reward mechanism would shift creators toward different content categories and how this shift would affect aggregate platform outcomes. Before deployment, however, it has no trajectory data under that new mechanism. The operational question is therefore how to predict the trajectory that would arise under a new payoff function using only trajectories observed under previous payoff functions.
Existing work provides important context but does not directly address this prediction problem. The population game dynamics literature primarily studies specified dynamics under a given payoff function, with central questions including convergence and stability [38]. Data-driven work has studied learning, identifying, and steering game dynamics from observed play [37, 6], while general dynamical-system learning methods learn vector fields or governing equations from trajectory data [7, 4]. These approaches can learn dynamics from observed trajectories, but they typically fit the realized motion under the observed payoff conditions. As a result, the learned dynamics may mix the payoff-dependent incentive with the population response structure, leaving no principled way to update the prediction when the payoff function changes.
To avoid this issue, we turn to structured population-game dynamics, where different dynamics encode different assumptions about how populations respond to payoffs [38]. Among these dynamics, Riemannian game dynamics (RmD) [32] provide a unified framework that includes commonly used models such as replicator dynamics [44] and projection dynamics [33]. RmD’s key idea is to model population motion as the outcome of balancing a payoff-dependent gain from motion against a payoff-independent cost from motion. This gain-cost view is natural in platform settings [49, 10, 9]: creators may be attracted to content categories with higher exposure or rewards, but switching categories can be costly. These adjustment costs can often be viewed as stable across payoff designs because they reflect practical barriers such as learning new production skills [11]. Motivated by this relative stability and following the RmD framework, we model the cost term as payoff-independent and use it as the shared population response structure for payoff-aware prediction. Once this cost term is learned from historical trajectories, a new payoff function can be plugged into the gain term to predict the induced population motion. We therefore propose PA-RmD, a payoff-aware Riemannian dynamics learning model that implements this gain-cost decomposition from data.
We analyze the proposed PA-RmD method both theoretically and empirically. On the theory side, we show that the learned model preserves the Nash equilibrium set of a fixed payoff function when all strategies retain positive population mass, and we prove finite-time one-step prediction bounds under a shared population response assumption. These bounds identify a payoff-coverage condition: prediction under a held-out payoff is controlled when its payoff direction is represented by the payoff directions observed in training. Empirically, we evaluate the framework on synthetic zero-sum games and potential games, as well as on a YouTube Trending multi-country dataset. The results show that the payoff-aware formulation can support prediction across held-out payoff designs.
To conclude, our main contributions are as follows:
We formulate payoff-aware prediction of population game dynamics: predicting population trajectories under held-out payoff designs from trajectories observed under previous payoff designs.
We instantiate this formulation with PA-RmD, a RmD-based framework that keeps payoff functions explicit and learns the shared population response structure from data.
We show that the learned model preserves fixed-payoff Nash equilibria and prove finite-time one-step prediction bounds under shared population response and payoff-coverage conditions.
We demonstrate the effectiveness of the framework on synthetic zero-sum games and potential games and a YouTube Trending multi-country dataset.
Related Work
Population games and dynamics. Population games provide a standard framework for modeling strategic interaction in large populations of anonymous agents through aggregate strategy shares and state-dependent payoffs [38]. This population-level viewpoint has been used in congestion and routing systems [33, 39], distributed control [35, 2], and evolutionary and biological systems [44, 20]. We adopt the same aggregate perspective because our goal is to predict how population behavior changes when the payoff environment changes.
Game dynamics specify how a population or a set of players adapts over time in response to payoff information. In population games, these dynamics are often formalized as revision protocols that convert current payoff vectors into feasible motion. Canonical examples include replicator, projection, logit, and Riemannian game dynamics [44, 33, 31, 32, 15]. More broadly, dynamics in games are also used as algorithmic and conceptual tools for studying decentralized adaptation and equilibrium computation [34, 19, 42, 12]. Prior work typically starts from a specified update rule and studies its stationarity, stability, convergence, regret, or ability to accelerate equilibrium computation [39, 13, 18]. In contrast, we treat the dynamic itself as the object to be learned from trajectory data, with the goal of predicting population-level evolution under held-out payoff environments.
Learning dynamical systems from data. Data-driven system identification learns dynamics from trajectories, usually as equations, vector fields, or transition maps. Sparse Identification of Nonlinear Dynamics discovers interpretable equations from candidate libraries [4], while neural ODEs parameterize continuous-time vector fields with neural networks [7]. Other methods incorporate structure for particle systems [5], fluid dynamics [26], social networks [46], and linear dynamical systems [17, 8, 48, 41]; multi-environment dynamics-learning methods study generalization across changing systems or latent environments [51, 22, 27].
Recent work has also learned dynamics or payoff information from game data. Lam et al. [24] estimate learning dynamics in routing games, and the SIAR framework [37] identifies game dynamics from scarce trajectories using polynomial vector fields with side-information constraints. SIAR-MPC further combines learned controlled dynamics with model predictive control to steer game dynamics toward desirable outcomes [6]. Relatedly, Bonafini et al. [3] recover individual payoff functions from observed evolutionary trajectories. These works infer dynamics, learning rules, or payoffs within an observed game environment. Our task is different: given a new payoff function and no trajectory under it, the model must predict aggregate evolution from a learned population response.
Preliminaries
Notations For a positive integer K, let [K]={1,…,K} and let 1 denote the all-ones vector in RK. We write ΔK={x∈R≥0K:1Tx=1} for the simplex, ΔK∘={x∈R>0K:1Tx=1} for its interior, and T={z∈RK:1Tz=0} for the tangent space of the simplex. For d≥1, S+d and S++d denote the sets of d×d symmetric positive semidefinite and positive definite matrices, respectively. We use ∥⋅∥2 for the Euclidean norm and ∥⋅∥op for the operator norm.
Population Games and Dynamics
We consider finite-strategy population games with strategy set [K]. A population state is a vector x∈ΔK, where xi denotes the mass, or share, of the population using strategy i. A population game is specified by a payoff function v:ΔK→RK, where vi(x) is the payoff received by an agent choosing strategy i when the aggregate population state is x. Let U denote the set of payoff functions.
A population dynamic specifies how the population state changes in response to the current payoff vector. On a finite time interval I=[0,T], we write the induced trajectory as
x˙t=F(xt,v(xt)),(1)
where F(xt,v(xt))∈T and T is the tangent space of the simplex. Thus, the dynamic converts the payoff vector at the current state into a feasible direction of motion for the population state.
We write X⊆ΔK for the reachable population-state domain generated by the dynamics under payoff functions v∈U and admissible initial conditions. Throughout the paper, we restrict attention to interior trajectories and assume that X⊆ΔK∘.
For a payoff function v:X→RK, define its Nash equilibrium set on the interior domain X as
N(v)={x∈X:vi(x)=vj(x),∀i,j∈[K]}.(2)
Indeed, the standard population-game Nash condition requires every strategy used with positive mass to attain maximal payoff. Since every strategy has positive mass on X⊆ΔK∘, this condition is equivalent to payoff indifference across all strategies.
Payoff-Aware Prediction of Population Game Dynamics
We observe population trajectories generated under multiple payoff functions. For each training instance n∈[N], a payoff function v(n)∈U induces a trajectory xt(n)∈X over a finite time interval [0,T]. The trajectories are observed at sampling interval δ>0, with observation grid I^δ={0,δ,2δ,…,⌊T/δ⌋δ}. The training data are a collection of observed trajectories, each containing the population states and the corresponding payoff vectors along the trajectory: Dobs={{(xt(n),v(n)(xt(n)))}t∈I^δ:n∈[N]}. For simplicity, we write the setup with a common observation grid across trajectories; the same formulation extends directly to trajectory-specific observation times.
At prediction time, we are given a payoff function w∈U and an initial state x0te∈X. The payoff function w may differ from those observed during training. The goal is to predict the population trajectory xtte induced by w from the initial state x0te. Thus the learning problem is payoff-aware: the payoff function is kept as an explicit input, and the model learns how population states respond to payoff vectors from the observed trajectories.
Example 3.1 (Online content platform). Consider an online content platform, such as TikTok or YouTube, on which creators choose among K content categories. The state xi is the share of creators producing category i content, and vi(x) is the expected payoff to producing category i when the platform-wide content distribution is x. Each training payoff function v(n) corresponds to a historical ranking, exposure, or reward rule, and the observed trajectory xt(n) records how creators reallocate across categories under that rule. At prediction time, w represents the platform rule used for forecasting. A payoff-aware model learns the shared adjustment structure from historical trajectories and combines it with w to predict the creator population trajectory from the initial state x0te.
Learning Payoff-Aware Riemannian Dynamics
Learning Procedure
To address the payoff-aware prediction problem defined in § 3.2, the model must explain how population motion changes when the payoff function changes. A model that only fits the realized motion in observed trajectories can mix the payoff-dependent incentive with the response rule of the population, leaving no clear way to update predictions under a new payoff function.
We therefore use the structure of population-game dynamics. Different dynamics encode different assumptions about how agents revise strategies in response to payoffs [38]; equivalently, the payoff function alone does not determine population motion, because the response rule is part of the model [34]. In our setting, the goal is to keep the payoff function explicit while learning the part of the population response that can be shared across payoff designs.
Payoff-aware Riemannian dynamics learning model. Riemannian game dynamics (RmD) [32] provide a natural formalism for this goal. We define a payoff-aware Riemannian dynamics learning model, denoted by PA-RmD. The model keeps the same structural form as RmD, but parameterizes the unknown shared cost-from-motion structure by Gθ:X→S++K and learns it from observed trajectories. Specifically, for any payoff function u∈U, the induced population dynamic is
x˙t=z∈Targmax{u(xt)Tz−21zTGθ(xt)z}.(3)
The parameter θ is shared across payoff functions, while the payoff function u remains an explicit input. Thus, PA-RmD does not learn a separate dynamic for each payoff function; it learns the payoff-independent adjustment structure that converts payoff incentives into population motion.
Interpretation. The variational problem in Equation (3) makes the gain-cost decomposition explicit. The linear term u(xt)Tz is the gain from motion: a direction z is favored when it increases the mass of strategies with high current payoffs and decreases the mass of strategies with low current payoffs. The quadratic term 21zTGθ(xt)z is the cost from motion: the matrix Gθ(xt) assigns a state-dependent cost to each feasible direction in T. In PA-RmD, this payoff-independent cost term is the shared population response structure to be learned from data. It can represent adjustment barriers, responsiveness to payoff differences, and constraints on reallocation. In the content-platform example, such costs capture how difficult it is for creators to shift toward a newly rewarded content category, for instance because doing so may require learning new production skills or rebuilding an audience.
Riemannian game dynamics also admit an individual-level interpretation through revision protocols, in which agents revise strategies according to payoff-dependent switching rates and the aggregate inflows and outflows induce the population dynamic. Under this interpretation, Gθ encodes how agents adjust behavior at the individual level, rather than merely a formal geometric object. We give the details in Appendix A.
Equivalent parametrization. The optimization problem in (3) admits an equivalent form that is easier to use for learning. Fix an orthonormal basis q1,…,qK−1 of T and let
U=[q1,…,qK−1]∈RK×(K−1).
Proposition 4.1. Fix U as above. The Riemannian model in (3) is equivalent to
x˙t=UCθ(xt)UTu(xt),(4)
where Cθ:X→S++K−1. Specifically, any metric Gθ:X→S++K induces this form with Cθ(x)=(UTGθ(x)U)−1. Conversely, any Cθ:X→S++K−1 can be represented by the Riemannian model in (3); one compatible choice is Gθ(x)=UCθ(x)−1UT+c11T/K, for any scalar c>0.
Thus, instead of learning the full metric Gθ, we learn Cθ directly. The parametrization in (4) is convenient because it replaces the constrained optimization over feasible simplex directions in (3) with a matrix-vector product.
Loss function. Let I^δ=I^δ∖{⌊T/δ⌋δ} denote the observation times that admit a one-step successor. We learn θ by matching the finite-difference approximation of the time derivative with the motion predicted by PA-RmD:
For the theoretical analysis, we impose a stronger interiority condition.
Assumption 4.1. Assume that X⊂ΔK∘ is a compact set.
This assumption focuses the analysis on regimes in which every strategy retains a non-zero population share. This interior-domain restriction is standard in population and evolutionary game dynamics, and is natural in large-population applications where each action or category can continue to attract some agents even when its population share is small [38, 30, 14].
Equilibrium preservation. We first provide a structural property of the model: for a fixed payoff function, the matrix C affects the transient motion but not the set of interior stationary states. Following the population-dynamic notation in (1), for a population dynamic F and payoff function u, define the stationary set induced by u as
References
[1]Sanjeev Arora, Elad Hazan, and Satyen Kale. The multiplicative weights update method: a meta-algorithm and applications. Theory of computing, 8(1):121–164, 2012.
[2]Julian Barreiro-Gomez, Nicanor Quijano, and Carlos Ocampo-Martinez. Constrained distributed optimization: A population dynamics approach. Automatica, 69:101–116, 2016.
[3]Mauro Bonafini, Massimo Fornasier, and Bernhard Schmitzer. Data-driven entropic spatially inhomogeneous evolutionary games. European Journal of Applied Mathematics, 34(1):106–159, 2023.arxiv.org/abs/2103.05429
[4]Steven L Brunton, Joshua L Proctor, and J Nathan Kutz. Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the national academy of sciences, 113(15):3932–3937, 2016.DOI
[5]Charlotte Bunne, Laetitia Papaxanthos, Andreas Krause, and Marco Cuturi. Proximal optimal transport modeling of population dynamics. In International Conference on Artificial Intelligence and Statistics, pages 6511–6528. PMLR, 2022.arxiv.org/abs/2106.06345
[6]Ilayda Canyakmaz, Iosif Sakos, Wayne Lin, Antonios V arvitsiotis, and Georgios Piliouras. Learning and steering game dynamics towards desirable outcomes. In 7th Annual Learning for Dynamics\ & Control Conference, pages 1512–1524. PMLR, 2025.arxiv.org/abs/2404.01066
[7]Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018.arxiv.org/abs/1806.07366
[8]Yanxi Chen and Vincent Poor. Learning mixture of linear dynamical systems. In International Conference on Machine Learning, 2022.arxiv.org/abs/2201.11211
[9]Sarah Dean, Evan Dong, Meena Jagadeesan, and Liu Leqi. Accounting for AI and users shaping one another: The role of mathematical models. Trans. Mach. Learn. Res., 2025, 2025.arxiv.org/abs/2404.12366
[10]Itay Eilat and Nir Rosenfeld. Performative recommendation: diversifying content via strategic incentives. In International Conference on Machine Learning, pages 9082–9103. PMLR, 2023.arxiv.org/abs/2302.04336
[11]Joseph Farrell and Paul Klemperer. Coordination and lock-in: Competition with switching costs and network effects. Handbook of industrial organization, 3:1967–2072, 2007.DOI
[12]Yoav Freund and Robert E Schapire. Adaptive game playing using multiplicative weights. Games and Economic Behavior, 29(1-2):79–103, 1999.DOI
[13]Drew Fudenberg and David K Levine. The theory of learning in games, volume 2. MIT press, 1998.
[14]Jacob K Goeree, Charles A Holt, and Thomas P Palfrey. Regular quantal response equilibrium. Experimental economics, 8(4):347–367, 2005.
[15]Saeed Hadikhahloo, Rida Laraki, Panayotis Mertikopoulos, and Sylvain Sorin. Learning in nonatomic games, part i: Finite action spaces and population games. Journal of Dynamics and Games, 9(4):433–460, 2022.
[16]Ernst Hairer, Gerhard Wanner, and Syvert P Nørsett. Solving ordinary differential equations I: Nonstiff problems. Springer, 1993.
[17]Moritz Hardt, Tengyu Ma, and Benjamin Recht. Gradient descent learns linear dynamical systems. Journal of Machine Learning Research, 2018.
[18]Keegan Harris, Ioannis Anagnostides, Gabriele Farina, Mikhail Khodak, Steven Wu, and Tuomas Sandholm. Meta-learning in games. In The Eleventh International Conference on Learning Representations, 2023.arxiv.org/abs/2209.14110
[19]Sergiu Hart and Andreu Mas-Colell. A simple adaptive procedure leading to correlated equilibrium. Econometrica, 68(5):1127–1150, 2000.
[20]Josef Hofbauer and Karl Sigmund. Evolutionary games and population dynamics. Cambridge university press, 1998.
[21]Pratik Kayal, Pascal Mettes, Nima Dehmamy, and Minsu Park. Large language models are natural video popularity predictors. In Findings of the association for computational linguistics: ACL 2025, pages 11432–11464, 2025.
[22]Matthieu Kirchmeyer, Yuan Yin, Jérémie Donà, Nicolas Baskiotis, Alain Rakotomamonjy, and Patrick Gallinari. Generalizing to new physical systems via context-informed dynamics model. In International conference on machine learning, pages 11283–11301. PMLR, 2022.arxiv.org/abs/2202.01889
[23]Haruka Kiyohara, Fan Yao, and Sarah Dean. Policy design for two-sided platforms with participation dynamics. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=qr4a4uS82y.arxiv.org/abs/2502.01792
[24]Kiet Lam, Walid Krichene, and Alexandre Bayen. On learning how players learn: estimation of learning dynamics in the routing game. In 2016 ACM/IEEE 7th International Conference on Cyber-Physical Systems (ICCPS), pages 1–10. IEEE, 2016.
[25]Siyang Li, Yize Chen, Yan Guo, Ming Huang, and Hui Xiong. Towards generalizable pde dynamics forecasting via physics-guided invariant learning. In International Conference on Learning Representations, volume 2026, pages 72050–72079, 2026.arxiv.org/abs/2509.24332
[26]Zongyi Li, Nikola Borislavov Kovachki, Kamyar Azizzadenesheli, Burigede liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021.
[27]Shixuan Liu, Yue He, Haotian Wang, Wenjing Yang, Yunfei Wang, Peng Cui, and Zhong Liu. Environment inference for learning generalizable dynamical system. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025.arxiv.org/abs/2510.19784
[28]Wei Min Loh, Sajib Kumer Sinha, Ankur Agarwal, and Pascal Poupart. A practical algorithm for feature-rich, non-stationary bandit problems. arXiv preprint arXiv:2603.16755, 2026.arxiv.org/abs/2603.16755
[29]Ryan McGrady, Kevin Zheng, and Ethan Zuckerman. One platform, four languages: Comparing english, spanish, hindi, and russian youtube. Social Media+ Society, 11(3):20563051251363216, 2025.DOI
[30]Richard D McKelvey and Thomas R Palfrey. Quantal response equilibria for normal form games. Games and economic behavior, 10(1):6–38, 1995.
[31]Panayotis Mertikopoulos and William H Sandholm. Learning in games via reinforcement and regularization. Mathematics of Operations Research, 41(4):1297–1324, 2016.arxiv.org/abs/1407.6267
[32]Panayotis Mertikopoulos and William H Sandholm. Riemannian game dynamics. Journal of Economic Theory, 177:315–364, 2018.arxiv.org/abs/1603.09173
[33]Anna Nagurney and Ding Zhang. Projected dynamical systems in the formulation, stability analysis, and computation of fixed-demand traffic network equilibria. Transportation Science, 31(2):147–158, 1997.DOI
[34]Christos Papadimitriou and Georgios Piliouras. Game dynamics as the meaning of a game. ACM SIGecom Exchanges, 16(2):53–63, 2019.DOI
[35]Nicanor Quijano, Carlos Ocampo-Martinez, Julian Barreiro-Gomez, German Obando, Andres Pantoja, and Eduardo Mojica-Nava. The role of population games and evolutionary dynamics in distributed control systems: The advantages of evolutionary game theory. IEEE Control Systems Magazine, 37(1):70–97, 2017.DOI
[36]Robert W Rosenthal. A class of games possessing pure-strategy nash equilibria. International journal of game theory, 2(1):65–67, 1973.DOI
[37]Iosif Sakos, Antonios Varvitsiotis, and Georgios Piliouras. Data-scarce identification of game dynamics via sum-of-squares optimization. In Web and Internet Economics LNCS 15534, page 651. Springer, 2024.arxiv.org/abs/2307.06640
[38]William H Sandholm. Population games and evolutionary dynamics. MIT press, 2010.
[39]William H Sandholm. Population games and deterministic evolutionary dynamics. In Handbook of game theory with economic applications, volume 4, pages 703–778. Elsevier, 2015.DOI
[40]Siavash Shahshahani. A new mathematical framework for the study of linkage and selection. American Mathematical Soc., 1979.DOI
[41]Max Simchowitz, Horia Mania, Stephen Tu, Michael I. Jordan, and Benjamin Recht. Learning without mixing: Towards a sharp analysis of linear system identification. In Conference on Learning Theory, 2018.arxiv.org/abs/1802.08334
[42]Vasilis Syrgkanis, Alekh Agarwal, Haipeng Luo, and Robert E Schapire. Fast convergence of regularized learning in games. Advances in Neural Information Processing Systems, 28, 2015.arxiv.org/abs/1507.00407
[43]Attila Szolnoki, Mauro Mobilia, Luo-Luo Jiang, Bartosz Szczesny, Alastair M Rucklidge, and Matjaž Perc. Cyclic dominance in evolutionary games: a review. Journal of the Royal Society Interface, 11(100), 2014.
[44]Peter D Taylor and Leo B Jonker. Evolutionary stable strategies and game dynamics. Mathematical biosciences, 40(1-2):145–156, 1978.DOI
[45]Samuele Tosatto, Joao Carvalho, Hany Abdulsamad, and Jan Peters. A nonparametric off-policy policy gradient. In International Conference on Artificial Intelligence and Statistics, pages 167–177. PMLR, 2020.arxiv.org/abs/2001.02435
[46]Rakshit Trivedi, Mehrdad Farajtabar, Prasenjeet Biswal, and Hongyuan Zha. Dyrep: Learning representations over dynamic graphs. In International conference on learning representations, 2019.
[47]Kang Wang, Renzhe Xu, and Bo Li. Lower bias, higher welfare: How creator competition reshapes bias-variance tradeoff in recommendation platforms? In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, pages 1435–1446, 2026.arxiv.org/abs/2511.20289
[48]Yuanyuan Wang, Biwei Huang, Wei Huang, Xi Geng, and Mingming Gong. Identifiability analysis of linear ode systems with hidden confounders. In Advances in Neural Information Processing Systems, 2024.arxiv.org/abs/2410.21917
[49]Renzhe Xu, Haotian Wang, Xingxuan Zhang, Bo Li, and Peng Cui. Competing for shareable arms in multi-player multi-armed bandits. In International Conference on Machine Learning, pages 38674–38706. PMLR, 2023.arxiv.org/abs/2305.19158
[50]Fan Yao, Chuanhao Li, Denis Nekipelov, Hongning Wang, and Haifeng Xu. How bad is top-k recommendation under competing content creators? In International conference on machine learning, pages 39674–39701. PMLR, 2023.arxiv.org/abs/2302.01971
[51]Yuan Yin, Ibrahim Ayed, Emmanuel de Bézenac, Nicolas Baskiotis, and Patrick Gallinari. Leads: Learning dynamical systems that generalize across environments. Advances in Neural Information Processing Systems, 34:7561–7573, 2021.arxiv.org/abs/2106.04546