Experiments and Considerations on Signal Control Using Deep Reinforcement Learning at Multiple Traffic Flow

Abstract

Expert engineers determined the parameters of signal control in traffic management systems. However, in recent years, the number of expert engineers is decreasing. Therefore, it is expected that AI substitution will save the workforce. Existing research on signal control using reinforcement learning compares conventional control methods with fixed traffic flow rates or for situations where there is a random inflow of vehicles. However, in reality, there are cases where traffic flows increase or decrease depending on the time of day or day of the week, and where the proportion changes. This paper compares the results of learning a signal control method using reinforcement learning, which is trained on a single traffic volume pattern, with existing control methods for multiple time-varying traffic flows for a single intersection. This experiment yields that reinforcement learning achieved more accurate signal control than the existing method in eight out of nine traffic flow patterns. The flexibility and versatility of reinforcement learning signal control were confirmed.