How can I approximately solve a 2-player zero-sum game by subselecting its rows/columns?

99 Views Asked by At

This is rather an open-ended question, but I'm posting here since I was not able to find a good resource elsewhere.

Say there's a two player zero sum game with payoff matrix $A$ that's $N \times H$. Call the value of this game $\lambda_N^*$. If $N$ is very large, we would be interested in finding $n\ll N$ somehow that might not do as well, but reasonably approximates $\lambda_N^*$. Let the value of the game of this subpayoff matrix be $\lambda_n^*$.

Here are my questions:

  1. Has there been any research on how to wisely select $n$, so that $\lambda_n^*$ best approximates $\lambda_N^*$?
  2. If so, can you provide a reference, and the main result, e.g. the approximation ratio, any assumptions necessary for the payoff matrix?
1

There are 1 best solutions below

6
On BEST ANSWER

Small support strategy profiles that achieve additively approximate Nash equilibria have been studied in both zero-sum games:

Lipton RJ, Young NE. Simple strategies for large zero-sum games with applications to complexity theory. InProceedings of the twenty-sixth annual ACM symposium on Theory of computing 1994 May 23 (pp. 734-740). ACM.

Althöfer I. On sparse approximations to randomized strategies and convex combinations. Linear Algebra and its Applications. 1994 Mar 1;199:339-55.

And, more recently, in general games:

Lipton RJ, Markakis E, Mehta A. Playing large games using simple strategies. InProceedings of the 4th ACM conference on Electronic commerce 2003 Jun 9 (pp. 36-41). ACM.

For an $n \times n$ game and a given $\epsilon$, one can find by brute force search over strategies that are uniform on multisets of pure strategies of size $\log n$ an approximate Nash equilibrium where no player can gain more than $\epsilon$ by deviating in time $n^{\frac{\log n}{\epsilon^2}}$.