String Stable Control of Connected Vehicles via Multi-Agent Lyapunov Actor-Critic

ICLR 2026 Conference Submission18578 Authors

Keywords: multi-agent reinforcement learning, string stability, connected vehicles, networked system control

Abstract: Networked or interconnected systems, such as urban transportation networks, rely on robust control strategies to ensure string stability-the concept that prevents the amplification of disturbances through the network. This capability is critical for system performance. A directly motivated example is the mitigation of phantom traffic jams, where string stability can suppress stop-and-go wave propagation. In this paper, we first establish a sufficient condition for scalable input-to-state stability (sISS), providing a theoretical guarantee for string stability. The derived conditions reveal a coupled shrinking relationship among the energy of different agents in the system, which depend on local Lyapunov functions but guarantee the global condition of sISS. Based on this theoretical foundation, we propose a practical and effective algorithm, named multi-agent Lyapunov actor-critic (MALAC), to achieve stable control in networked systems. Numerical simulations demonstrate that MALAC can ensure the string stability in the cooperative adaptive cruise control task.

Primary Area: reinforcement learning

Submission Number: 18578