[Paper Review] Traffic regulation via controlled speed limit

Maria Laura Delle Monache, Benedetto Piccoli+1 more|arXiv (Cornell University)|Mar 15, 2016

Traffic control and management21 references3 citations

TL;DR

This paper formulates and solves an optimal control problem for traffic regulation using variable speed limits on a single road, modeled via the LWR equation with a Newell-Daganzo flux function. It derives analytical expressions for cost functional variations using needle-like control perturbations and compares three numerical strategies—instantaneous policy, random exploration, and gradient descent—showing that gradient descent achieves within 10% of the best performance at 15% of the computational cost.

ABSTRACT

We study an optimal control problem for traffic regulation via variable speed limit. The traffic flow dynamics is described with the Lighthill-Whitham-Richards (LWR) model with Newell-Daganzo flux function. We aim at minimizing the $L^2$ quadratic error to a desired outflow, given an inflow on a single road. We first provide existence of a minimizer and compute analytically the cost functional variations due to needle-like variation in the control policy. Then, we compare three strategies: instantaneous policy; random exploration of control space; steepest descent using numerical expression of gradient. We show that the gradient technique is able to achieve a cost within 10% of random exploration minimum with better computational performances.

Motivation & Objective

• To address the challenge of minimizing L² error between actual and desired outflow in traffic flow via variable speed limits.
• To establish existence of a minimizer for the optimal control problem under bounded variation constraints on the speed limit policy.
• To develop and compare numerical strategies for solving the complex, delayed input-output mapping inherent in traffic control.
• To evaluate the trade-off between control performance and computational cost in real-time traffic regulation.
• To provide a mathematically rigorous framework for gradient-based optimization in hyperbolic conservation laws with delayed effects.

Proposed method

• Models traffic dynamics using the Lighthill-Whitham-Richards (LWR) PDE with Newell-Daganzo flux function under free-flow assumptions.
• Defines the optimal control problem as minimizing the L² norm of the difference between achieved and desired outflow.
• Derives analytical expressions for one-sided variations of the cost functional using needle-like perturbations in the control policy.
• Utilizes distributional derivatives and BV function theory to characterize cost sensitivity, enabling gradient computation.
• Implements three numerical strategies: instantaneous policy (based on real-time density), random exploration (binary tree search over control space), and gradient descent using computed gradients.
• Applies the link entering time (LET) concept to model time delays in control effects, linking control history to future outflow.

Experimental results

Research questions

• RQ1 Can a minimizer exist for the optimal speed limit control problem under bounded variation constraints on the control policy?
• RQ2 How can the sensitivity of the cost functional to small control perturbations be analytically characterized in the presence of time delays?
• RQ3 What is the relative performance of gradient descent compared to random exploration and instantaneous policy in minimizing outflow error?
• RQ4 How does computational cost scale across different control strategies while maintaining acceptable control accuracy?
• RQ5 Can gradient-based methods achieve near-optimal performance with feasible computational load for real-time implementation?

Key findings

• The cost functional is Lipschitz continuous in the space of bounded variation functions, ensuring existence of a minimizer.
• Analytical expressions for cost variations due to needle-like control perturbations are derived, involving integrals with respect to the distributional derivative of the solution.
• The gradient descent method achieves a cost within 10% of the best result from random exploration, demonstrating strong performance efficiency.
• Gradient descent requires approximately 15% of the computational cost of random exploration, making it computationally viable for real-time applications.
• The instantaneous policy performs poorly compared to random exploration but has negligible computational cost, limiting its practicality despite low overhead.
• Random exploration yields the best performance but produces control policies with high total variation, which may be impractical for real-world deployment.