Incremental approximate saddle-point computation in zero-sum matrix games

We consider the problem of approximately and efficiently computing saddle-point values for zero-sum matrix games. This problem arises in scenarios where the game's exact value is hard to compute, either because the columns of the matrix are revealed incrementally in time, or because the game's strategy space is too large for traditional methods (e.g., linear programming) to be effective in practice. We lever-age the established adaptive multiplicative weights algorithm but introduce a novel simple criterion to determine whether the minimizer's best strategy needs to be approximately re-computed as a new column of the matrix is introduced. Our main results are two-fold. First, we show that our proposed incremental approach achieves the same accuracy as applying the adaptive multiplicative weights algorithm on the entire matrix, if known a priori. Secondly, we argue that our approach can be computationally more efficient than simply re-computing the minimizer's best strategy upon addition of every new column of the matrix. Specifically, for the case when the columns of the matrix are generated independently and from the same distribution, we characterize the probability that the expected number of times the best response is re-computed exceeds a given fraction of the total number of columns in the matrix. Numerical simulations indicate even more significant computational improvement as compared to the analytic result.