Voice of Panda Science Nature cover: Humans lose to AI again, this time playing “GT Racing

Nature cover: Humans lose to AI again, this time playing “GT Racing

Nature cover: Humans lose to AI again, this time playing “GT Racing” Article | Academic headlines, author | Cooper, editor and reviewer | Kou Jianchao Many potential applications of artificial intelligence (AI) involve making more optimized real-time decisions when interacting with humans, and competitive or gaming games are the best stage to showcase. Today, a…

Nature cover: Humans lose to AI again, this time playing “GT Racing”

Article | Academic headlines, author | Cooper, editor and reviewer | Kou Jianchao

Many potential applications of artificial intelligence (AI) involve making more optimized real-time decisions when interacting with humans, and competitive or gaming games are the best stage to showcase.

Today, a cover article published in the journal Nature reports that AI has beaten world-champion-level human players in the racing battle game Gran Turismo (GT racing). The AI ​​program, called “Gran Turismo (GT) Sophy,” is a neural network driver that demonstrates exceptional speed, handling, and driving strategy while adhering to the rules of racing.

The core team that completes the development of this AI program is from Sony AI Division (Sony AI). The “GT Racing” series of games is developed by Polyphony Digital in Japan, which faithfully reproduces the nonlinear control challenges of real racing cars and encapsulates complex multi-agents. Interactive, the game is released on game console platforms such as Sony PlayStation and PSP. It is a popular racing game with a very realistic manipulation experience.

If there is the blessing of this AI program, it is estimated that human players will no longer be able to run the enhanced version of the stand-alone program, right?

Figure | Game screenshots (Source: GT Racing)

The researchers believe the achievement could make racing games more interesting and provide high-level competitions for training professional drivers and discovering new racing skills. The approach is also expected to be used in real-world systems such as robots, drones, and self-driving cars.

Fast and furious on the track

Driving a racing car requires great skill. Modern Formula 1 cars display astonishing engineering precision, however, the sport’s popularity has less to do with the car’s performance PK than with the skill and grit of top drivers in pushing the car’s performance to the limit related. Success on the track has been fast and furious for over a century.

Figure | Formula 1 racing competition (Source: GNEWS)

The goal of a car race is simple: if you finish the track in less time than your competitors, you win. Achieving this, however, requires extremely complex physical warfare, and navigating the track requires careful use of friction between the tires and the road, which is limited.

To win the race, drivers must choose a trajectory that keeps the car within changing friction limits. Brake too early on a turn and your car will slow down and waste time. Brake too late, and as you approach the tightest part of the turn, you won’t have enough cornering power to maintain your desired course trajectory. Braking too hard may cause the body to spin.

 

As a result, professional racers are very good at finding and maintaining the limits of their car lap after lap throughout the race.

Although the handling limits of racing cars are complex, they are well described physically, so it is only a matter of course that they can be calculated or learned.

In recent years, Deep Reinforcement Learning (DRL) has become a key component of AI research milestones in fields such as Atari, StarCraft, and Dota. For AI to have an impact on robotics and automation, researchers must demonstrate the ability to successfully control complex physical systems. In addition, many potential applications of AI technology require interactions in close proximity to humans, while respecting imprecise human norms. Automotive Competitions are a typical area of ​​these challenges.

Figure | Game data comparison (Source: Nature)

In recent years, research into autonomous racing has accelerated using full-scale, large-scale, and simulated vehicles. A common approach is to precompute trajectories and use model predictive control to execute those trajectories. However, small modeling errors can be catastrophic when driving at the absolute limit of friction.

Competing with other drivers places higher demands on AI modeling accuracy and introduces complex aerodynamic interactions, further prompting engineers to improve control schemes to continuously predict and adapt to the optimal trajectory of the track. Driving a car off the track to compete with a human driver is not empty talk.

The making of “AI racer”

During the development of GT Sophy, researchers explored various ways of using machine learning to avoid modeling complexity, including using supervised learning to model vehicle dynamics and using imitation learning, evolutionary methods, or reinforcement learning to learn to drive Strategy.

To be successful, a racer must be highly skilled in four areas: (1) racing control, (2) racing tactics, (3) racing etiquette, and (4) racing strategy.

In order to control the car, drivers have detailed knowledge of their vehicle dynamics and the characteristics of the track. On this basis, the driver builds the tactical skills needed to execute precise maneuvers by defending the opponent. At the same time, drivers must obey highly refined but imprecise rules of sportsmanship, and finally, drivers use strategic thinking when simulating opponents and deciding when and how to attempt to overtake.

Simulation racing is a field that requires real-time, continuous control in a highly realistic, complex physics environment, and the success of GT Sophy in this environment shows for the first time that it is possible to train better-than-best-in-class racing cars across a range of car and track types. A better AI agent for human racers.

This result can be seen as another important step in the continued development of computers in competitive tasks such as chess, Go, adventure, poker, and StarCraft.

Figure | Training of GT Sophy (Source: Nature)

Notably, GT Sophy learned to take a detour in just a few hours and outperformed 95% of the human runners in the dataset, it trained for another nine days and accumulated over 45,000 hours of driving and lap time A tenth of a second was reduced until the lap times stopped improving.

Progress rewards alone are not enough to motivate AI programs to win games. If the human opponent is fast enough, the AI ​​program will learn to follow and try to accumulate more rewards to overtake without risking a potentially catastrophic collision.

To evaluate the GT Sophy, the researchers pitted the GT Sophy against top GT drivers in two events. The GT Sophy achieved superhuman timing performance on all three tracks tested. It was able to execute several types of turns, effectively Take advantage of drifts, disrupt vehicles behind, intercept opponents and perform other emergency maneuvers.

While the GT Sophy demonstrates adequate tactical skills, there are still many areas to improve, especially when it comes to strategic decision-making. For example, the GT Sophy will sometimes leave enough space on the same track to give opponents an opportunity.

Figure | AI drivers surpass human players (Source: Nature)

More worthy of attention outside of competitive games

Regarding e-sports and gaming games, it is not uncommon for AI to defeat human beings, and it is certain that AI will become stronger and stronger. There is not much suspense and significance. The key is to see how these superhuman AI programs can effectively overcome industrial bottlenecks and truly benefit human life.

On February 10, 1996, the supercomputer Deep Blue first challenged the world chess champion Kasparov and lost 2:4. Challenged again in May 1997, and in the end Deep Blue defeated Kasparov 3.5:2.5, becoming the first computer system to beat a world chess champion within the standard time limit.

But Deep Blue’s flaw is that it has no intuition and does not have a real “smart soul”. It can only rely on super computing power to make up for the shortcomings of analytical thinking. Deep Blue, who won the game, soon retired.

 

In March 2016, Google AI’s AlphaGo defeated the Go world champion Lee Sedol in four games, which is considered a real milestone in AI. AlphaGo used a combination of Monte Carlo tree search and two deep neural networks. , Under this design, the computer can spontaneously learn to analyze and train like the human brain, and continuously learn to improve chess.

Since then, various AI program rookies have emerged one after another. On December 10, 2018, DeepMind’s artificial intelligence AlphaStar developed for the real-time strategy game StarCraft can completely abuse 99.8% of the world’s human professional players.

Undoubtedly, the current GT Sophy is another continuation of an AI victory.

J.Christian Gerdes, a professor of mechanical engineering at Stanford University, believes that the impact of the GT Sophy Institute may go far beyond video games. How much neural networks should be used, and how much should be based solely on physics, deserves further exploration.

Overall, neural networks are the undisputed champions when it comes to perceiving and recognizing objects in the surrounding environment. Trajectory planning, however, is still the domain of physics and optimization, and GT Sophy’s success on gaming tracks suggests that neural networks may one day play a bigger role in the software of automated vehicles than they do today.

More challenging may be the change per lap. In real life, the tire condition of the car changes between laps, and the human driver has to adapt to this change throughout the race. Can GT Sophy do the same thing with more data? Where does this data come from? This will give artificial intelligence more room for evolution.

https://www.tmtpost.com/6003921.html

作者: wanfeng

联系我们

联系我们

0898-88881688

在线咨询: QQ交谈

邮箱: email@wangzhan.com

工作时间:周一至周五,9:00-17:30,节假日休息

关注微信
微信扫一扫关注我们

微信扫一扫关注我们

关注微博
返回顶部