Traffic Impact Analysis of a Deep Reinforcement Learning-based Multi-lane Freeway Vehicle Control

Reinforcement learning is one of the methods that has been used to realize optimal driving. Most studies have focused on evaluating learning performance of a fraction of vehicles controlled by reinforcement learning. It is unclear how these controlled vehicles influence other vehicles. We conducted several experiments examining the impact of multiple vehicles controlled by reinforcement learning on traffic flow. The simulations were performed on a three-lane freeway with lane regulation at the end of one of the lanes. The controlled vehicles were trained to drive as fast as possible and run non-cooperatively. We found out that controlled vehicles could run faster than human-driven vehicles. Moreover, we anticipated that if multiple vehicles were run selfishly, it would adversely affect traffic flow. Contrary to expectations, the experimental results showed that even if numerous controlled vehicles drive selfishly, the negative impact on overall traffic would be small.