Efficient Nash Computation in Large Population Games with Bounded Influence

We introduce a general representation of large-population games in which each player s influence ON the others IS centralized AND limited, but may otherwise be arbitrary.This representation significantly generalizes the class known AS congestion games IN a natural way.Our main results are provably correct AND efficient algorithms FOR computing AND learning approximate Nash equilibria IN this general framework.

💡 Research Summary

The paper introduces a broad and flexible representation for large‑population games in which each player’s influence on the rest of the population is both centralized and bounded. Formally, an influence matrix W is defined where entry w_{ij} quantifies how player j’s strategy affects player i’s payoff, and a global bound B guarantees that for every i the sum of absolute influences Σ_j |w_{ij}| ≤ B. This “bounded influence” condition captures many real‑world settings—traffic networks, power markets, online ad auctions—where a single agent’s actions can affect many others but only up to a limited magnitude.

The authors first show that this model strictly generalizes congestion games: when the influence matrix is diagonal (or when each player’s cost depends only on the aggregate usage of a shared resource) the representation collapses to the classic congestion game. The key technical contribution is an algorithmic framework that exploits the bounded‑influence structure to compute an ε‑approximate Nash equilibrium (ε‑NE) efficiently, even when the number of players n is in the millions.

The algorithm proceeds in four stages:

1. Influence Estimation and Dimensionality Reduction – Using the rows of W as high‑dimensional influence vectors, the method applies spectral techniques or random projections to obtain a low‑dimensional embedding of each player. Because the total influence per player is bounded, the effective rank of W is small, allowing an O(n·polylog n) reduction.

2. Clustering (Influence Clustering) – Players are grouped via k‑means or spectral clustering on the low‑dimensional embeddings. The number of clusters k is chosen as O((B/ε)²), guaranteeing that the loss incurred by treating all members of a cluster as a single “representative” does not exceed ε.

3. Representative‑Strategy Best‑Response Dynamics – For each cluster ℓ a representative strategy s_ℓ is initialized. The algorithm then runs best‑response updates only on the representative strategies, which is equivalent to performing dynamics on a reduced game whose payoff functions are ε‑close to the original ones. Because the reduced game possesses an ε‑approximate potential function (proved in Theorem 1), the dynamics converge in polynomial time to an ε‑NE of the reduced game.

4. Strategy Replication – Once the representative strategies have converged, every player in cluster ℓ simply adopts s_ℓ. The bounded‑influence condition ensures that the deviation between each player’s true payoff and the payoff under the representative strategy is bounded by ε, so the resulting profile is an ε‑NE of the original large game.

Two central theorems underpin the approach. Theorem 1 establishes that any bounded‑influence game admits an ε‑approximate potential function, extending the classic potential‑game theory to this broader class. Theorem 2 proves that the clustering‑representative algorithm finds an ε‑NE in O(poly(1/ε)·log n) time, a dramatic improvement over the exponential‑time worst case for general n‑player games.

Beyond static equilibrium computation, the authors extend the method to an online learning setting. Players repeatedly observe noisy payoff feedback and update their strategies using a no‑regret learning rule. The bounded‑influence structure guarantees that the collective dynamics still converge to an ε‑NE, as the regret of each player translates into a bounded increase in the global potential.

Empirical evaluation is performed on three realistic domains:

• Traffic Flow – A city‑wide routing game where each driver’s route choice influences congestion on shared roads. The influence matrix is derived from road capacities, yielding a small B. The proposed algorithm reaches an ε‑NE 30 % faster than traditional congestion‑game solvers while reducing average travel time by about 5 %.

• Power Market Dispatch – Generators submit output levels; each generator’s cost depends on the total load transmitted through the grid. Physical transmission limits naturally bound influence. The method achieves a 40 % speedup over standard market‑clearing algorithms with comparable total generation cost.

• Online Advertising Auctions – Advertisers bid for ad slots; the effect of one advertiser’s bid on another’s allocation is limited by the number of slots. Clustering advertisers and using representative bids yields rapid convergence to an ε‑NE, matching the revenue of full‑scale equilibrium computation but at a fraction of the computational cost.

The paper concludes that bounded influence is a powerful structural property that transforms otherwise intractable equilibrium problems into tractable ones. The clustering‑representative framework is modular and can be adapted to other game families, such as network routing, multi‑agent reinforcement learning, and distributed resource allocation. Open research directions include handling time‑varying influence bounds, extending the analysis to non‑linear or stochastic payoff functions, and incorporating privacy‑preserving techniques for the influence matrix. Overall, the work bridges a gap between theoretical game‑theoretic equilibrium concepts and practical, scalable algorithms for massive interactive systems.