Bilinear Games: Polynomial Time Algorithms for Rank Based Subclasses

Motivated by the sequence form formulation of Koller et al. [19], this paper defines bilinear games, and proposes efficient algorithms for its rank based subclasses. Bilinear games are two-player non-cooperative single-shot games with compact polytopal strategy sets and two payoff matrices (A, B) such that when (x, y) is the played strategy profile, the payoffs of the players are xTAy and xTBy respectively. We show that bilinear games are very general and capture many interesting classes of games like bimatrix games, two player Bayesian games, polymatrix games, two-player extensive form games with perfect recall etc. as special cases, and hence are hard to solve in general. Existence of a (symmetric) Nash equilibrium for (symmetric) bilinear games follow directly from the known results. For a given bilinear game, we define its Best Response Polytopes (BRPs) and characterize the set of Nash equilibria as fully-labeled pairs in the BRPs.We consider a rank based hierarchy of bilinear games, where rank of a game (A, B) is defined as rank(A+ B). In this paper, we give polynomial time algorithms to compute Nash equilibrium for special classes of bilinear games: – Rank-1 games (i.e., rank(A + B) = 1). – FPTAS for constant rank games (i.e., rank(A+ B) is constant). – When rank(A) or rank(B) is constant. This improves the results by Lipton et al. [22] and Kannan et al. [17], for bimatrix games with low rank matrices.