[Paper Review] Linearly Solvable Mean-Field Traffic Routing Games

Takashi Tanaka, Ehsan Nekouei+2 more|arXiv (Cornell University)|Mar 4, 2019

Transportation Planning and Optimization41 references4 citations

TL;DR

This paper introduces a linearly solvable mean-field traffic routing game where drivers face congestion costs affine in the logarithm of route usage. By leveraging a backward-only linear system solution, it achieves exact mean-field equilibrium computation with strong time-consistency, offering a computationally efficient alternative to standard forward-backward HJB-FPK methods in mean-field games.

ABSTRACT

We consider a dynamic traffic routing game over an urban road network involving a large number of drivers in which each driver selecting a particular route is subject to a penalty that is affine in the logarithm of the number of drivers selecting the same route. We show that the mean-field approximation of such a game leads to the so-called linearly solvable Markov decision process, implying that its mean-field equilibrium (MFE) can be found simply by solving a finite-dimensional linear system backward in time. Based on this backward-only characterization, it is further shown that the obtained MFE has the notable property of strong time-consistency. A connection between the obtained MFE and a particular class of fictitious play is also discussed.

Motivation & Objective

• To model large-population dynamic traffic routing as a mean-field game with congestion costs affine in the logarithm of route usage.
• To demonstrate that this class of games admits a linearly solvable structure, enabling efficient equilibrium computation.
• To establish strong time-consistency of the mean-field equilibrium through backward-only computation.
• To connect the derived equilibrium to fictitious play dynamics in large-population settings.

Proposed method

• Formulates a discrete-time stochastic dynamic game where drivers select routes based on randomized policies over a network.
• Models the cost function as an affine function of the logarithm of the number of drivers on the same route, representing a congestion penalty.
• Applies mean-field approximation to decouple individual driver dynamics while preserving cost coupling.
• Shows that the mean-field equilibrium (MFE) satisfies a linearly solvable Markov decision process (LS-MDP), solvable via backward matrix multiplications only.
• Derives a backward-only characterization that ensures strong time-consistency of the MFE.
• Establishes a connection between the MFE and a specific class of fictitious play dynamics through asymptotic analysis.

Experimental results

Research questions

• RQ1 Can a mean-field traffic routing game with logarithmic congestion costs be solved efficiently using linear solvability?
• RQ2 Does the backward-only solution of the LS-MDP framework yield a time-consistent mean-field equilibrium?
• RQ3 How does the derived MFE relate to fictitious play in large-population settings?
• RQ4 What is the limiting behavior of the N-player game as the number of drivers tends to infinity?
• RQ5 Can the mean-field equilibrium be characterized without solving a coupled forward-backward HJB-FPK system?

Key findings

• The mean-field equilibrium is computed by solving a finite-dimensional linear system backward in time, eliminating the need for iterative forward-backward solvers.
• The solution exhibits strong time-consistency, a property stronger than standard MFEs derived from HJB-FPK systems.
• The equilibrium is shown to be the limit of symmetric Nash equilibria in the finite-N game, with convergence rate δ_N → 0.
• A unique symmetric equilibrium exists in the finite-N single-stage game, proven via KKT conditions and intermediate value theorem on monotonic functions.
• The MFE is asymptotically equivalent to a fictitious play process, linking it to learning dynamics in large populations.
• The computational advantage is significant: only backward matrix multiplications are required, enabling scalable computation in large networks.