GT Sophy, the AI ​​that dominates real drivers on Gran Turismo

Which game will resist artificial intelligence? After go, chess or even Starcraft II, it is now the turn of Gran Turismo players to bow to an AI. A few days before the launch of the 7^and edition of the famous simulation automotive by Polyphony Digital, Sony unveiled GT Sophyan artificial intelligence capable of beating the best pilots!

In July, this AI had first beaten the best humans in races against the clock, that is to say, it was alone on the track. But, in October, a milestone was reached since GT Sophy beat human players on a real race with overtaking on the track, but also strategy. Precisely, it is in this area that AI impresses.

Unprecedented driving

” AI rolls in ways we never imagined “, underlines Takuma Miyazono, one of the world references for this video game. Same observation at Kazunori Yamauchi, the creator of Gran Turismo and general manager of the studio, Polyphony Digital, which gives the example of braking in full curve.

” Typically, racing drivers learn to brake in a straight line with the goal of slowing down in the curve to accelerate out of the corner. GT Sophy does not necessarily do this. When it enters a curve, it actually brakes as it enters the curve. Usually, when entering a curve, the load is only on the two front tires; but with GT Sophy you have the load distributed over three tires, two in the front and one in the rear as well. This allows the car to brake while it is turning. »

What distinguishes this AI from others is its type of learning. We thus knew the“deep” learning ”, and it is already part of our daily life today in research on the Internet or the fight against spam. The AI ​​is trained with millions of examples, and then it is able to be autonomous in finding similar images or weeding out spam in our emails. GT Sophy was entitled to it with more than 45,000 hours of learning, based on years of games and stored on a thousand playstation !

The power of reinforcement learning

To complete this machine learning which is very crude, Sony has opted for reinforcement learning (reinforcement learning). It is a type of machine learning used to train AIs to make decisions in an environment with a system of rewards or penalties for each action depending on the results they lead to. This method applied to sports simulation is so relevant and cutting-edge that it is entitled to an article this week and even the cover of the prestigious magazine Nature.

The diagram below shows how an AI interacts with its environment. She takes an action in the world, receives a reward (or penalty) and an updated description of the state of the world to determine her next action. Applied to the car race, it is a question of reacting to the maneuvers of the adversaries, but also to the modifications of the track. The difficulty was becoming aware of the unwritten rules of motor racing, such as avoiding collisions and not cutting corners in dangerous ways.

To challenge GT Sophy, you will unfortunately have to wait since she will not be included in Gran Turismo 7corn via a later update. According to experienced pilots who have tested it, its strong point lies in its ability to have a “human” driving style. At no time did they feel like they were challenging a computer whose piloting is usually very mechanical and predictable.