University of Toronto Technology Overview
There are two separate inventions that use machine learning to improve traffic flow:
1. MIND: Multi-modal Intelligent Deep Traffic Signal Control
MIND is an intelligent traffic signal controller (software) that comprises three major parts:
Convolutional Neural Networks for interpreting high dimensional sensory data;
Fully Connected Neural Networks as function approximator for managing continuous features of the traffic network; and
Reinforcement Learning as the brain of the controller for learning how to optimize the travel time for users of the traffic network.
Each of these parts are trained together to achieve a single goal, optimizing the traffic signal, with consideration for efficiently moving the most people.
2. MARLIN: Multi-Agent Reinforcement Learning For Integrated Network of Adaptive Traffic Signal Controllers
MARLIN is a system and method for determining and generating the optimal traffic signal action based on reinforcement learning (RL) and the state and actions of linked traffic signals that also use RL. Rewards are based on the optimal control policy for the entire system.
In most state of the art research in this area, traffic signal controllers are designed based on assumption of perfect detection and hence encounter challenges when applied in the field in real-life intersections. They mostly work with queue length information, assuming to be seamlessly and perfectly provided by cameras, while in practice queue detection using image processing methods faces many problems including limited detection area, inaccurate detection, and weather-related detection problems.
On top of this problem, some researchers attempt to include partial information from upstream cars joining the queues, in order to provide more information for the traffic signal controllers. Such information needs to be heavily pre-processed, the procedure is usually case-specific, and it includes changing the structure of the controller.
Another problem with the state-of-the-art traffic signal controller is their limited ability to handle multiple modes of transportation (i.e., low occupancy passenger vehicles vs. high occupancy transit vehicles). There are 3 major issues in this area:
Providing priority for transit always causes interruption for regular traffic and in most cases it leads to higher average delays over all the modes,
Introducing a new mode needs expert knowledge to extract useful information for the controller,
and it results in more complicated state-space for already high-dimensional state-space of the controller.