Algorithms for non-convexS _n -games

Abstract

We considerS-games in which the setS is the parametric curve{(p ₁ (t),⋯, p _n(t)): t ∈ [0, 1]} and thep _i(t) are real polynomials. These games will be referred to asS _n -games. Two iterative algorithms are given in the case ofn = 2. One gives linear convergence to an optimal of player I, and the other gives monotone convergence. In the case of arbitraryn we give an algorithm of the cutting plane family converging to the value. The principal features of this algorithm are that the hyperplanes arise intrinsically as tangents to an associated concave function which is not in general differentiable, and the linear subproblems arise as matrix games. Because of the latter property inactive constraints are in principle automatically dropped. We give a game theoretic convergence proof. This algorithm may be used multiply for bounding optimals of player I.