[Paper Review] Learning Sparse Polymatrix Games in Polynomial Time and Sample Complexity

Asish Ghoshal, Jean Honorio|arXiv (Cornell University)|Jun 18, 2017

Game Theory and Applications3 references3 citations

TL;DR

This paper proposes an $ε$-Nash equilibrium recovery method for sparse polymatrix games using $ε_{1,2}$-group regularized logistic regression, achieving polynomial time and sample complexity of $\mathcal{O}(m^4d^4\log(pd))$. It proves information-theoretic optimality and exact Nash equilibrium recovery under separability conditions, validated via synthetic and real-world data including Supreme Court, congressional, and U.N. voting records.

ABSTRACT

We consider the problem of learning sparse polymatrix games from observations of strategic interactions. We show that a polynomial time method based on $\ell_{1,2}$-group regularized logistic regression recovers a game, whose Nash equilibria are the $ε$-Nash equilibria of the game from which the data was generated (true game), in $\mathcal{O}(m^4 d^4 \log (pd))$ samples of strategy profiles --- where $m$ is the maximum number of pure strategies of a player, $p$ is the number of players, and $d$ is the maximum degree of the game graph. Under slightly more stringent separability conditions on the payoff matrices of the true game, we show that our method learns a game with the exact same Nash equilibria as the true game. We also show that $Ω(d \log (pm))$ samples are necessary for any method to consistently recover a game, with the same Nash-equilibria as the true game, from observations of strategic interactions. We verify our theoretical results through simulation experiments.

Motivation & Objective

• To develop a polynomial-time algorithm for learning sparse polymatrix games from observed strategic interactions.
• To establish sample complexity bounds for consistent recovery of Nash equilibria in sparse polymatrix games.
• To prove information-theoretic lower bounds on sample requirements for any consistent recovery method.
• To validate the method on synthetic data and real-world voting datasets (Supreme Court, Congress, U.N. General Assembly).
• To recover game structure and Nash equilibria that reflect real-world coalition patterns in voting behavior.

Proposed method

• Uses $\ell_{1,2}$-group regularized logistic regression to learn the payoff matrices of sparse polymatrix games from observed strategy profiles.
• Models the game as a graphical game where each player's payoff depends on interactions with neighbors in a graph of maximum degree $d$.
• Treats the problem as a classification task: distinguishing Nash equilibrium profiles from non-equilibrium profiles using logistic regression with group sparsity.
• Applies a group-lasso penalty over player-specific payoff matrices to enforce sparsity in the game graph structure.
• Employs a modified observation model allowing zero measure for non-Nash profiles, generalizing prior work.
• Uses empirical risk minimization with $\ell_{1,2}$ regularization to recover a game whose Nash equilibria approximate those of the true game.

Experimental results

Research questions

• RQ1 Can a polynomial-time method recover $\epsilon$-Nash equilibria of a sparse polymatrix game from observed strategy profiles?
• RQ2 What is the minimum number of samples required to consistently recover the Nash equilibrium set of a sparse polymatrix game?
• RQ3 Under what conditions can the method recover the exact Nash equilibrium set of the true game?
• RQ4 Is the proposed method information-theoretically optimal in terms of sample complexity?
• RQ5 Can the method uncover meaningful coalition structures in real-world voting data?

Key findings

• The proposed $\ell_{1,2}$-regularized logistic regression method recovers an $\epsilon$-Nash equilibrium of the true game with high probability using $\mathcal{O}(m^4d^4\log(pd))$ samples.
• Under separability conditions on the payoff matrices, the method recovers the exact same Nash equilibria as the true game.
• The method achieves information-theoretic optimality, as $\Omega(d\log(pm))$ samples are necessary for consistent recovery of the Nash equilibrium set.
• Simulation results show a phase transition in PSNE recovery success, matching the theoretical sample complexity bound.
• On real-world data, the method recovers meaningful connected components corresponding to liberal/conservative blocs in the Supreme Court, party blocs in Congress, and regional voting coalitions in the U.N.
• The price of anarchy (PoA) computed from recovered games ranges from 1.6 to 3.0 across datasets, quantifying inefficiency due to selfish behavior.