Interceptor agents in traffic flow: A reinforcement learning approach

This paper presents a simulation-based study of autonomous vehicle agents trained via reinforcement learning to intercept a moving target in a multi-lane traffic flow. The traffic environment is modeled using a stochastic cellular automaton based on the revised S-NFS model, and interceptor agents are implemented as neural networks trained with the APPO (Advantage-Weighted Proximal Policy Optimization) algorithm. Two types of reinforcement learning policies are considered: a single-agent strategy and a cooperative multi-agent strategy involving two interceptors. The interception process is divided into two phases: (i) convergence — reaching the target in dense traffic, and (ii) stopping — reducing the target’s speed through coordinated behavior. The performance of RL-based agents is evaluated against several heuristic baselines across a wide range of traffic densities. Key metrics include convergence time, average speed, lane-changing frequency, and impact on local traffic flow and density. Results indicate that RL-trained interceptors achieve faster convergence and more adaptive behavior compared to heuristic strategies. Moreover, while a single agent is generally insufficient to fully stop the target, coordinated multi-agent policies reliably achieve immobilization. The findings highlight the potential of reinforcement learning for real-time multi-agent decision-making in dynamic traffic environments.