Jam or Flow?
In the United States, traffic jam has cost about $305 billion in 2017. Each person spends average 60 hours of commuting time per year. Ten years ago, a Japanese researcher conducted an experiment, by asking 20 drivers to drive around a ring at 30 kph. The experiment showed that traffic was initially smooth and orderly, but within half a minute, the traffic waves were formed and some cars had to stop in a jam. Last year, some other researchers conducted a new experiment by adding a single autonomous vehicle in the ring, which showed autonomous vehicles played a stable role in traffic flows and kept a smooth traffic due to lack of randomly accelerating or decelerating.
Based on the experiment results, researchers from the Department of Energy’s Lawrence Berkeley National Laboratory (Berkeley Lab) decided to launch a project associated with deep reinforcement learning (RL) to train autonomous vehicles for the goal of improving the traffic flow. This project called Congestion Impact Reduction via CAV-in-the-loop Lagrangian Energy Smoothing (CIRCLES) is based on a software framework — Flow.
Flow is an innovative framework that constructs and solves various RL tasks on traffic, and discovers schemes for evaluating and optimizing traffic flows by using an advanced open-source micro-simulator. Flow can simulate thousands of human-driving or autonomous vehicles in custom traffic scenarios.
There are still some limitations of technical breakthroughs both on autonomous vehicles and optimization of urban infrastructure. However, in the promising future, we can assume that this framework could be adopted in many cities with ring roads or highways, and will help establish a new city system when autonomous vehicles are widely adopted.
Source:
Kheterpal, N.& Parvate, K.& Wu, C.& Kreidieh, A.&Vinitsky, E.& Bayen,A. M.(2018). Flow: Deep Reinforcement Learning for Control in SUMO
