Game on Random Environment, Mean-field Langevin System and Neural Networks
This work addresses theoretical challenges in game theory and machine learning, with applications to neural network optimization, but appears incremental as it builds on existing mean-field and Langevin frameworks.
The paper tackles the problem of analyzing games regularized by relative entropy with strategies coupled through a random environment, proving existence and uniqueness of equilibria and convergence of mean-field Langevin systems to these equilibria. It applies these results to analyze stochastic gradient descent in deep neural networks for supervised learning and generative adversarial networks.
In this paper we study a type of games regularized by the relative entropy, where the players' strategies are coupled through a random environment variable. Besides the existence and the uniqueness of equilibria of such games, we prove that the marginal laws of the corresponding mean-field Langevin systems can converge towards the games' equilibria in different settings. As applications, the dynamic games can be treated as games on a random environment when one treats the time horizon as the environment. In practice, our results can be applied to analysing the stochastic gradient descent algorithm for deep neural networks in the context of supervised learning as well as for the generative adversarial networks.