Being stuck in traffic is frustrating and expensive. Beyond headaches and missed appointments, traffic congestion costs U.S. drivers some $300 billion annually.
Researchers suggest self-driving cars — even in small numbers — will dramatically improve traffic flow. Lex Fridman and his team at MIT created a game to accelerate this future.
DeepTraffic simulates a typical highway environment, and its players control their own car using deep learning. The simulation makes complex technical concepts accessible for beginners, and the gamification pushes experts to develop completely new techniques.
Gaming Traffic with Neural Nets
Imagine you’re driving on a busy freeway in Los Angeles. You must decide how closely to follow the car in front of you, when to change lanes and how to avoid hitting other cars while you navigate. This is called path planning. With DeepTraffic, anyone can design and train a deep neural network to do it.
During a session at the GPU Technology Conference in May in Silicon Valley, Fridman talked about how the game relies on reinforcement learning. This is an approach to achieving AI where a neural network is rewarded for taking the desired action. By repeating these rewards over and over, the network learns how to perform.
DeepTraffic was originally created for a class Fridman teaches at MIT. It gained popularity when both the course content and the game were opened to the public. With over 12,000 submissions to date, DeepTraffic is competitive. The top-ranked users appear on a leaderboard alongside the fastest speed their network achieved.
The competitive aspect of the game increases the fun factor, but the real world stakes are much higher. Autonomous vehicles must plan safe paths from one point to another. AI is required given the complexity of driving tasks. Educational tools like DeepTraffic help train the next generation of AI developers, and surface solutions that will transform the automotive ecosystem.
Fridman’s complete GTC talk is available below. Watch it to learn more about the hierarchical levels of path planning, the pros and cons of reinforcement learning, and technical details about training a network for DeepTraffic.
Feature image by rust.bucket, licensed via Creative Commons.