A Globally Convergent Flow for Time-Dependent Mean Field Games and a Solver-Agnostic Framework for Inverse Problems
Hanwei Yan, Xianjin Yang, Jingguo Zhang
TL;DR
A solver-agnostic framework is proposed, in which parameter updates are computed by implicitly differentiating the discrete MFG equations satisfied by the converged MFG solution, rather than by differentiating through a particular forward solver.
Abstract
Mean field games (MFGs) model the limit of large populations of strategically interacting agents, yet both forward and inverse problems remain challenging. For the forward problem, a difficulty is to design numerical methods with global convergence guarantees whose convergence does not depend on careful initialization. For the inverse problem, a difficulty is to decouple parameter optimization from the forward solver, so that parameter updates do not depend on implementation details, and the inverse method does not need to be reformulated when the forward solver is changed. We address both issues as follows. For the forward problem, we propose a monotone Hessian-Riemannian flow for time-dependent MFGs on the feasible manifold of densities. The flow preserves the positivity of densities and is proved to be globally convergent. For the inverse problem, we cast parameter estimation as an outer optimization problem over the unknown coefficients, with the MFG system solved in an inner step for each parameter value. For solving this problem, we consider an adjoint-based gradient method together with a Gauss-Newton acceleration. This leads to a solver-agnostic framework, in which parameter updates are computed by implicitly differentiating the discrete MFG equations satisfied by the converged MFG solution, rather than by differentiating through a particular forward solver. We demonstrate the approach on several stationary and time-dependent MFG examples, where the Gauss-Newton method consistently requires fewer outer iterations than gradient descent.