Technology: AI can outrace champion Gran Turismo drivers
A paper reporting on an artificial intelligence (AI) agent who has won against a world champion-level player in the car racing game "Gran Turismo", which is a close battle with rival cars,Nature Will be published in.The agent demonstrated outstanding speed, vehicle control and tactics while preserving race etiquette.This finding may be useful for basic research on autonomous navigation and AI.
Many of the applicable examples of AI require real-time decision making in the physical system while interacting with humans.Auto racing is an example of such a situation.The driver must perform complex tactical operations to overtake or block rival vehicles while manipulating the vehicle at the static friction limit.Race simulations such as "Gran Turismo," one of the PlayStation games, recreates the challenges of controlling a real race car, making it a difficult application for machine learning.
Now, Peter Wurman and colleagues have taught an AI agent named Gran Turismo Sophy (GT Sophy) how to play "Gran Turismo" using deep reinforcement learning. GT Sophy was trained to master the techniques of efficient acceleration and braking used on the race course and learned how to find alternative paths under various circumstances or when blocked by an opponent.One of the most difficult aspects of training high-performance AI is to ensure that you avoid the penalties associated with violating race etiquette, a set of loose rules set by outside human referees. GT Sophy played a close battle with four of the world's best esports drivers in a situation where different race challenges were presented depending on the combination of three types of vehicles (including vehicles exceeding 1 km / h) and car race tracks. I won.
Computers continue to defeat the strongest humans in competitive challenges (chess, poker, etc.), and this result is a new step in taking over this trend. Wurman et al. Say that this finding could not only improve the enjoyment of car racing games, but also create a realistic and high-level competitive situation for training professional drivers and discovering new racing techniques. Suggests.This technique may also be applicable to real-world systems (robots, aerial drones, self-driving cars, etc.).
[Original English »]
"Highlights of Nature Magazines" is a translation of a release made by Nature's public relations department for the press.If you need more accurate and detailed information, be sure to read the original paper.
* This article is reprinted from "Nature Japan Featured Highlights".
Reprinted from: "Technology: AI overtakes Gran Turismo champion driver'
