First-order algorithm with O(ln(1/ε )) convergence for ε -equilibrium in two-person zero-sum games

  • Authors:

  • Andrew Gilpin;Javier Peña;Thomas Sandholm

  • Affiliations:

  • Computer Science Department, Carnegie Mellon University, Pittsburgh, PA;Tepper School of Business, Carnegie Mellon University, Pittsburgh, PA;Computer Science Department, Carnegie Mellon University, Pittsburgh, PA

  • Venue:
  • AAAI'08 Proceedings of the 23rd national conference on Artificial intelligence - Volume 1
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

We propose an iterated version of Nesterov's first-order smoothing method for the two-person zero-sum game equilibrium problem minx∈Q1 maxy∈Q2 xT Ay = maxy∈Q2 minx∈Q1 xTAy. This formulation applies to matrix games as well as sequential games. Our new algorithmic scheme computes an Ε-equilibrium to this min-max problem in O(κ(A) In(1/Ε)) first-order iterations, where κ(A) is a certain condition measure of the matrix A. This improves upon the previous first-order methods which required O(1/Ε) iterations, and it matches the iteration complexity bound of interior-point methods in terms of the algorithm's dependence on Ε. Unlike the interior-point methods that are inapplicable to large games due to their memory requirements, our algorithm retains the small memory requirements of prior first-order methods. Our scheme supplements Nesterov's algorithm with an outer loop that lowers the target Ε between iterations (this target affects the amount of smoothing in the inner loop). We find it surprising that such a simple modification yields an exponential speed improvement. Finally, computational experiments both in matrix games and sequential games show that a significant speed improvement is obtained in practice as well, and the relative speed improvement increases with the desired accuracy (as suggested by the complexity bounds).