Eco-cooperative adaptive cruise control for platoons in mixed traffic using single-agent and multi-agent reinforcement learning

Signalized arterials create stop-and-go traffic, leading to collisions, delays, wasted energy, and discomfort. Connected Automated Vehicles (CAVs), using Cooperative Adaptive Cruise Control (CACC), can potentially mitigate these issues by optimizing speeds with shared information. However, the traffic environment in CACC research on signalized roads is predominantly generated through simulations. This paper compares various eco-friendly CACC methods based on reinforcement learning (RL) for CAVs operating with Human-driven Vehicles (HVs) on signalized arterials. Methods analyzed include Deep Deterministic Policy Gradient (DDPG), Soft Actor-Critic (SAC), and their multi-agent versions (MADDPG, MASAC), trained and tested on naturalistic data from the pNEUMA dataset. These RL methods are benchmarked against human-driven trajectories and the Intelligent Driver Model (IDM) in mixed platoon scenarios. Results show that DDPG and SAC excel in vehicle performance (safety, efficiency, energy, comfort), while MADDPG and MASAC perform best in platoon stability. Key factors influencing performance include platoon characteristics, vehicle position, and preceding vehicle type.