By Yossi Doctor, Adi Borochov, Nimrod Alon, Nadav Alon
We aim to design an AI traffic lights controller to optimize and ensure better traffic flow using optimal light switching behavior.
We will use two approaches: reinforcement learning and search.
Given a rush-hour traffic stream, the main goal of the AI traffic lights controller is to let all the cars pass as quickly as possible while avoiding collisions. With reinforcement learning, the agent can be rewarded and penalized based on predetermined rules, such as priority of passage for emergency vehicles, making it learn to solve the problem. With search, the agent can be pointed using heuristics toward the final state of no traffic. These two approaches allow the AI controller to calculate the optimal switching behavior for the traffic lights to shorten waiting times at junctions and reduce journey times.
Using a traffic flow simulator, we will measure the average waiting and journey times and compare them at different stages of the learning process.