Stability Analysis in Mean-Field Games via an Evans Function Approach

Author(s):  
Piyush Grover

This work is concerned with stability analysis of stationary and time-varying equilibria in a class of mean-field games that relate to multi-agent control problems of flocking and swarming. The mean-field game framework is a non-cooperative model of distributed optimal control in large populations, and characterizes the optimal control for a representative agent in Nash-equilibrium with the population. A mean-field game model is described by a coupled PDE system of forward-in-time Fokker-Planck (FP) equation for density of agents, and a backward-in-time Hamilton-Jacobi-Bellman (HJB) equation for control. The linear stability analysis of fixed points of these equations typically proceeds via numerical computation of spectrum of the linearized MFG operator. We explore the Evans function approach that provides a geometric alternative to solving the characteristic equation.

Symmetry ◽  
2020 ◽  
Vol 13 (1) ◽  
pp. 14
Author(s):  
Luca Di Persio ◽  
Matteo Garbelli

We provide a rigorous mathematical formulation of Deep Learning (DL) methodologies through an in-depth analysis of the learning procedures characterizing Neural Network (NN) models within the theoretical frameworks of Stochastic Optimal Control (SOC) and Mean-Field Games (MFGs). In particular, we show how the supervised learning approach can be translated in terms of a (stochastic) mean-field optimal control problem by applying the Hamilton–Jacobi–Bellman (HJB) approach and the mean-field Pontryagin maximum principle. Our contribution sheds new light on a possible theoretical connection between mean-field problems and DL, melting heterogeneous approaches and reporting the state-of-the-art within such fields to show how the latter different perspectives can be indeed fruitfully unified.


2019 ◽  
Vol 29 (08) ◽  
pp. 1413-1464 ◽  
Author(s):  
Guilherme Mazanti ◽  
Filippo Santambrogio

This paper considers a mean field game model inspired by crowd motion where agents want to leave a given bounded domain through a part of its boundary in minimal time. Each agent is free to move in any direction, but their maximal speed is bounded in terms of the average density of agents around their position in order to take into account congestion phenomena. After a preliminary study of the corresponding minimal-time optimal control problem, we formulate the mean field game in a Lagrangian setting and prove existence of Lagrangian equilibria using a fixed point strategy. We provide a further study of equilibria under the assumption that agents may leave the domain through the whole boundary, in which case equilibria are described through a system of a continuity equation on the distribution of agents coupled with a Hamilton–Jacobi equation on the value function of the optimal control problem solved by each agent. This is possible thanks to the semiconcavity of the value function, which follows from some further regularity properties of optimal trajectories obtained through Pontryagin Maximum Principle. Simulations illustrate the behavior of equilibria in some particular situations.


2020 ◽  
Vol 26 ◽  
pp. 33
Author(s):  
Yurii Averboukh

In the paper, we examine the dependence of the solution of the deterministic mean field game on the initial distribution of players. The main object of study is the mapping which assigns to the initial time and the initial distribution of players the set of expected rewards of the representative player corresponding to solutions of mean field game. This mapping can be regarded as a value multifunction. We obtain the sufficient condition for a multifunction to be a value multifunction. It states that if a multifunction is viable with respect to the dynamics generated by the original mean field game, then it is a value multifunction. Furthermore, the infinitesimal variant of this condition is derived.


2020 ◽  
Vol 9 (4) ◽  
Author(s):  
Thibault Bonnemain ◽  
Thierry Gobron ◽  
Denis Ullmo

Mean Field Games provide a powerful framework to analyze the dynamics of a large number of controlled agents in interaction. Here we consider such systems when the interactions between agents result in a negative coordination and analyze the behavior of the associated system of coupled PDEs using the now well established correspondence with the non linear Schrödinger equation. We focus on the long optimization time limit and on configurations such that the game we consider goes through different regimes in which the relative importance of disorder, interactions between agents and external potential vary, which makes possible to get insights on the role of the forward-backward structure of the Mean Field Game equations in relation with the way these various regimes are connected.


Author(s):  
Sarah Perrin ◽  
Mathieu Laurière ◽  
Julien Pérolat ◽  
Matthieu Geist ◽  
Romuald Élie ◽  
...  

We present a method enabling a large number of agents to learn how to flock. This problem has drawn a lot of interest but requires many structural assumptions and is tractable only in small dimensions. We phrase this problem as a Mean Field Game (MFG), where each individual chooses its own acceleration depending on the population behavior. Combining Deep Reinforcement Learning (RL) and Normalizing Flows (NF), we obtain a tractable solution requiring only very weak assumptions. Our algorithm finds a Nash Equilibrium and the agents adapt their velocity to match the neighboring flock’s average one. We use Fictitious Play and alternate: (1) computing an approximate best response with Deep RL, and (2) estimating the next population distribution with NF. We show numerically that our algorithm can learn multi-group or high-dimensional flocking with obstacles.


Author(s):  
Виктория Сергеевна Корниенко ◽  
Владимир Викторович Шайдуров ◽  
Евгения Дмитриевна Карепова

Представлен конечно-разностный аналог дифференциальной задачи, сформулированной в терминах теории “игр среднего поля” (mean field games). Задачи оптимизации такого типа формулируются как связанные системы параболических дифференциальных уравнений в частных производных типа Фоккера - Планка и Гамильтона - Якоби - Беллмана. Предложенный конечно-разностный аналог обладает основными свойствами оптимизационной дифференциальной задачи непосредственно на дискретном уровне. В итоге он может служить как приближение, сходящееся к исходной дифференциальной задаче при стремлении шагов дискретизации к нулю, так и как самостоятельная оптимизационная задача с конечным числом участников. Для предложенного аналога построен алгоритм монотонной минимизации функционала стоимости, проиллюстрированный на модельной экономической задаче In most forecasting problems, overstating or understating forecast leads to various losses. Traditionally, in the theory of “mean field games”, the functional responsible for the costs of implementing the interaction of the continuum of agents between each other is supposed to be dependent on the squared function of control of the system. Since additional external factors can influence the player’s strategy, the control function of a dynamic system is more complex. Therefore, the purpose of this article is to develop a computational algorithm applicable for more general set of control functions. As a research method, a computational experiment and proof of the stability of the constructed computational scheme are used in this study. As a result, the numerical algorithm was applied on the problem of economic interaction in the presence of alternative resources. We consider the model, in which a continuum of consumer agents consists of households deciding on heating, having a choice between the cost of installing and maintaining the thermal insulation or the additional cost of electricity. In the framework of the problem, the convergence of the method is numerically demonstrated. Conclusions. The article considers a model of the strategic interaction of continuum of agents, the interaction of which is determined by a coupled differential equations, namely, the Fokker - Planck and the Hamilton - Jacobi - Bellman one. To approximate the differential problem, difference schemes with a semi-Lagrangian approximation are used, which give a direct rule for minimizing the cost functional


Sign in / Sign up

Export Citation Format

Share Document