Neural Information Processing Systems http://nips.cc/

It can also be seen as a local approximation of the nonconvex-nonconcave minimax games--e.g., the generative adversarial network (GAN)

We study a wide class of non-convex non-concave min-max games that generalizes over standard bilinear zero-sum games.

Stochastic min-max optimization has gained interest in the machine learning community with the advancements in GANs and adversarial training. Although game optimization is fairly well understood in the deterministic setting, some issues persist in the stochastic regime. Recent work has shown that stochastic gradient descent-ascent methods such as the optimistic gradient are highly sensitive to noise or can fail to converge. Although alternative strategies exist, they can be prohibitively expensive. We introduce Omega, a method with optimistic-like updates that mitigates the impact of noise by incorporating an EMA of historic gradients in its update rule. We also explore a variation of this algorithm that incorporates momentum. Although we do not provide convergence guarantees, our experiments on stochastic games show that Omega outperforms the optimistic gradient method when applied to linear players.

The extragradient method has recently gained increasing attention, due to its convergence behavior on smooth games. In $n$-player differentiable games, the eigenvalues of the Jacobian of the vector field are distributed on the complex plane. Thus, compared to classical (i.e., single player) minimization, games exhibit more convoluted dynamics, where the extragradient method succeeds while simple gradient method could fail. Yet, in this work, instead of focusing on a specific problem class, we follow a reverse path: starting from the momentum extragradient method as the selected optimizer, and using polynomial-based analyses, we identify problem subclasses where the use of momentum in extragradient motions lead to optimal performance. Based on the hyperparameter setup, we show that the extragradient with momentum exhibits three different modes of convergence: when the eigenvalues are distributed $i)$ on the real line, $ii)$ both on the real line along with complex conjugates, and $iii)$ only as complex conjugates. We then derive the optimal hyperparameters for each case, and show that it achieves an accelerated convergence rate.