Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.06292

Learning controls in mean field games: the fictitious play and related gradient descent

Charles Meynard
View original source ↗Read source document ↗Export metadata ↓

Abstract

We investigate two learning procedures and their relationship to the formation of equilibrium in potential mean field games (MFGs). The first one is a version of the fictitious play in which players observe directly the actions (controls) of other players and learn the distribution of the MFG through an update rules on controls. This leads to a new iterative procedure which converges for a wide class of potential MFGs of controls, even in the presence of common noise. On the other hand we also consider the learning procedure of a principal agent trying to solve a mean field control problem with only local gradient information. We prove that under very general assumption, the gradient descent of such agent converges to a solution of the associated potential MFG.

Explore related subjects

math.OCmath.PR

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Charles Meynard. 2026-10-05. Learning controls in mean field games: the fictitious play and related gradient descent. https://arxiv.org/abs/2610.06292

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Diagonally forced systems and the spectral signature of matrix cycles

Given any square matrix, $\mathbf{M}$, whose diagonal elements are negative, and which are multiplied by a variable, $σ$, we wish to find the minimal $σ$ such that the eigenvalue of $\mathbf{M}_σ$ is exactly zero. By Gershgorin, we know that $\mathbf{M}_σ$ can be made stable by making $σ$ large enough. We prove a relation which analytically determines when and how we are able to find the value of $σ$ such that the maximal eigenvalue is exactly zero. In so doing, we prove the equivalence of the roots of the characteristic polynomial of $\mathbf{M}_σ$ and the eigenvalues that arise from a scaling operation on $\mathbf{M}$. Further, through the characteristic polynomial, we are able to isolate the dominant feedback cycles comprising the elements of the matrix which, under the action of $σ$, ensures the stability of the system. We then explore, using the stabilising and destabilising cycles within the coefficients of the characteristic polynomial, an intrinsic spectral signature associated with any square matrix based on the sign (or zero) of its respective elements.

math.OC↗
arXiv2026

Nonlinear Control Allocation: A Polynomial Approach

Control allocation maps the virtual control inputs requested by a high-level control law onto the physical actuators of a system. When the actuator authority is strongly nonlinear, existing methods rely on local linearization, on incremental updates, or on general nonlinear programming, and may be inaccurate, slow, or trapped in local optima. This paper introduces Polynomial Control Allocation, in which the cost and constraints of the allocation problem are polynomial or rational functions of the states and inputs; radical, trigonometric and tabulated nonlinearities are accommodated by lifting. The first order optimality conditions then form a parametric system of polynomial equations. Its Gr{ö}bner basis and the associated multiplication matrices are computed once, offline. Online, the matrices are evaluated at the current state and all solutions are recovered from the eigenvectors of a random linear combination of them, so that the globally optimal allocation is selected at a predictable computational cost. The framework is illustrated on thrust control of a variable-pitch propeller and on multicopter force allocation, and its computational cost is quantified in terms of polynomial evaluation and eigenvector computation. The method is validated experimentally on a Quanser helicopter model at a 500 Hz control rate, where a two-layer allocation scheme combined with linear model predictive control decouples travel and elevation over multiple full rotations. Monte Carlo comparisons with nonlinear programming and incremental nonlinear control allocation show that Polynomial Control Allocation achieves the lowest allocation error together with the shortest computation time.

math.OC↗
arXiv2026

Notes on the discretization of TV-NORM regularized inverse potential problems

We describe a method to discretize optimization problems arising in the regularization of linear inverse problem having compact forward operator defined on 3-D valued measures, compactly supported on a fixed set. The criterion is a quadratic residual attached to the data, with an additive penalization of the total variation of the measure.

math.OC↗