Investigation of Multi-agent reinforcement learning on merge ramp for avoiding car crash on highway

Car crashing on the entrance of highway becomes the major concern of traffic control system because the merge ramp area can be considered as the bottleneck of the highway. When considering the autonomous driving vehicles, the multi-agent reinforcement learning (MARL) is one of the most popular techniques which is widely studied by researchers and tried to implement in practice. Previously, the performance of MARL were evaluated without concerning the driving law such as length of prohibiting area and length of entrance area. Therefore, in this paper, we take the as length of prohibiting area and length of entrance area into the account to evaluate the performance of MARL. As a result, the MARL can avoid the car crashing effectively while achieve high rewards and high vehicle speed. When the traffic becomes congested, MARL cannot avoid the crashing because there is no available area for entering the highway.