Multi-Agent Deep Reinforcement Learning for Traffic optimization through Multiple Road Intersections using Live Camera Feed.

Traffic signals provide one of the primary means to administer conflicting traffic flows. Existing signal control strategies, operating on hand-crafted rules, fail to efficiently, autonomously adapt to the changing traffic patterns. Each signal control system independently manages one intersection at a time and regulates navigation of vehicles through that intersection. Current systems cannot co-operate to optimize aggregate traffic flows through multiple road intersections. Consequently, they are susceptible to making myopic signal control decisions that might be effective locally, but not globally. Instead, we propose a system of multiple, coordinating traffic signal control systems. This paper presents the first application of multi-agent deep reinforcement learning (DRL) to achieve traffic optimization through multiple road intersections solely based on raw pixel input from CCTV cameras in real time. This set of traffic control agents is shown to significantly outperform independently operating (both DRL-trained and loop-induced) adaptive signal control systems, by increasing traffic throughput and reducing the average time a vehicle spends in an intersection. Additionally, this paper, introduces attention-based visualization to interpret and validate the proposed multi-agent signal control methodology.