Optimising Traffic Signal Control in the Colombo area through Reinforcement Learning

Abstract

"Traffic congestion is a major problem in Sri Lanka. It has been found that this has caused many negative effects on various aspects of the country. One of the main reasons for this is fixed-time phase-based traffic control, which has proven to be ineffective. It has been found that the majority of daily commuters in Colombo use motorcycles and scooters. Existing research provides evidence that these vehicles are affected more by adverse weather conditions, which leads to higher RTAs while affecting the mental and physical well-being of the drivers. After a thorough analysis of the problem, it is proposed to prioritise two-wheelers while optimising overall traffic. Which is a challenging problem. This entire scenario can be framed as MDP; therefore, to address the problem, a MARL-based ATSC solution was developed with the SUMO platform, where the Nugegoda area is simulated with unique Sri Lankan traffic dynamics. Each intersection in the simulation is controlled by a RL agent, where PPO has been used as the algorithm for the agents. The RL agent at the intersection is supposed to control the traffic based on the observations while trying to maximise its reward. This study proposes customised observation spaces and reward functions for the agents to address the above-mentioned problem. The developed model was trained for over 300,000 timesteps on the simulation platform, and agents have made significant progress over the timesteps. The trained model is evaluated against a fixed-time phase-based simulation in the same identical environment. The proposed model performs much better compared to the fixed-time-phase based method while optimising traffic and prioritising two-wheelers. According to the results obtained, the MARL-based method drops the mean waiting time of normal traffic and two-wheeler traffic by 91.2% and 93.5%, respectively, compared to the traditional fixed-time-phase based method."