Combined extreme learning machine and max pressure algorithms for traffic signal control

Nowadays, rush-hour traffic congestion problems persist in most major cities around the world, resulting in increased pollution, noise, and stress for citizens. Therefore, an optimal traffic light strategy is needed. For this purpose, several models have been proposed. However, these models often overlook the non-stationarity of traffic, which occurs due to changing traffic conditions over time. Additionally, these models are steady-state process models, leading to a decrease in their predictive power over time. To address these issues, this paper proposes the combination of two algorithms: a passive Extreme Learning Machine with periodic mini-batch learning (PB-ELM) for predicting traffic flow and the Max Pressure control algorithm (MPA) for signal control. In the first step, the passive periodic Extreme Learning Machine (PB-ELM) adjusts quickly and regularly based on new data, overcoming traffic non-stationarity and improving long-term performance. In the second step, the MPA is preferred for signal control due to its simplicity and speed. The PB-ELM-MPA model is a combination of predictive algorithms that takes the current road network conditions as input and predicts the flow of vehicles at intersections. The model utilizes learned characteristics of the source and destination roads to estimate the number of vehicles in each movement. The PB-ELM outputs serve as the starting point for the max-pressure algorithm, which reduces congestion by considering only the vehicles on road segments closest to the intersection and selecting the highest pressure at each time interval. The proposed PB-ELM-MPA model is evaluated on an isolated intersection simulated with the SUMO micro-simulator, demonstrating a significant improvement in avoiding traffic jams. The total staying time of all vehicles present at the intersection is reduced by 65% compared to the fixed configuration of traffic lights. Additionally, CO2 emissions and fuel consumption are reduced by approximately 34% compared to the classic MPA and Deep Q-Network approaches.