This paper considers a class of nonsmooth nonconvex-nonconcave min-max problems in machine learning and games. We first provide sufficient conditions for the existence of global minimax points and local minimax points. Next, we establish the first-order and second-order optimality conditions for local minimax points by using directional derivatives. These conditions reduce to smooth min-max problems with Fréchet derivatives. We apply our theoretical results to generative adversarial networks (GANs) in which two neural networks contest with each other in a game. Examples are used to illustrate applications of the new theory for training GANs.
1 January 2019 in Journal of Computational Mathematics
The Journal of Computational Mathematics published bi-monthly. It is an international journal covering all branches of modern computational mathematics such as numerical linear algebra, numerical optimization, computational geometry, numerical PDEs and inverse problems. Papers containing new ideas, creative approaches and/or innovative applications as well as invited reviews are expected to appear regularly in the journal. JCM is a refereed international journal sponsored by the Institute of Computational Mathematics and Scientific/Engineering Computing (ICMSEC) of the Chinese Academy of Sciences.
We present a versatile method that enumerates all or a user-specified number of longest sensitisable paths in the whole circuit or through specific components. The path information can be used for design and test of circuits affected by statistical process variations. The algorithm encodes all aspects of the path search as an instance of the Boolean Satisfiability Problem (SAT), which allows the method not only to benefit from recent advances in SAT-solving technology, but also to avoid some of the drawbacks of previous structural approaches. Experimental results for academic and industrial benchmark circuits demonstrate the method's accuracy and scalability.
21 March 2020 in Computational Optimization and Applications
In this paper, we consider stochastic optimization problems with $k$th order dominance constraints induced by full random recourse. We first introduce the optimization model with dominance constraints and its distributionally robust counterpart. After establishing the closedness and convexity of the feasible solution set, we derive the qualitative stability about the constraint set mapping with respect to the simultaneous perturbations of the random variable and the benchmark. Furthermore, we obtain the quantitative stability results of the constraint set mapping, the optimal value function and the optimal solution set under the Hausdorff metric, which extend the present results to the locally Lipschitz continuity case. Finally, for stochastic optimization problems with $k$th order distributionally robust dominance constraints, we derive the corresponding quantitative stability results under the appropriate pseudometric.
10 July 2021 in Journal of Optimization Theory and Applications