HumanLight

"AI can do a better job at traffic management"

Paper: HumanLight: Incentivizing Ridesharing via Human-centric Deep Reinforcement Learning in Traffic Signal Control

Researchers at University of California at Berkeley (among other contributors) are using machine learning to optimize traffic patterns.

Mainly, they wanted to use algorithms to create more time saving traffic signals, while keeping in mind single passengers and HOV interactions at intersections.

Source: HumanLight arxiv paper, Dimitris M. V. et al

• Firstly they defined the problem formally with an optimization for max pressure

  • Max Pressure or MP is defined as the difference between the total queue length on incoming approaches and outgoing approaches

• Then they generated intersections scenarios and introduced active agents

  • Active Agent are vehicles that are in range of an intersection within the action interval window.

• Finally, they used Reinforcement Learning (Deep Q Learning) to train models and compared their algorithms with current SOTA controllers

  • Reinforcement Learning allows researchers to train their models by defining a “reward” state for their active agents to stimulate certain behavior

  • They optimized on vehicle and people metrics for traffic time, queue length and delay time

HumanLight displays strong performance across all considered road network configurations. As per analysis, HumanLight is designed to equitably allocate green times at intersections. By rewarding HOV riders with more green times, HumanLight achieves reduced person delays and queues.

The more people use public vehicles, the smoother the traffic becomes.

So Essentially,

"AI can do a better job at traffic management"