War stories: Forza learned to love neural nets to train AI drivers

Produced by Justin Wolfson, edited by Shander Garrison. Click here for the transcript.

Once uppermost, Forza The franchise is now firmly established within the Pantheon of great racing games. The first installment was made as a response to the Xbox Gran turismo, But also with a healthy helping of online multiplayer racing. Since then, it has evolved with Microsoft’s Xbox consoles with more realistic graphics and sometimes more precise physics in track-focused Forza motorsport Along with the series, the open-world adventure (and even a journey to the Lego dimension) is evolving Forza Horizon sport.

If you are one of the millions of people who have played the Forza racing game, then you probably know about the AI ​​opponents of the game, called Drivers. When the first driveatars started in Forza motorsport In 2005, they were greatly improving on NPCs racing in other driving games, often following the same preprogrammed route around the track. “It was a machine-learning system on a hard drive that used the Bayesian neural network to record [racing] Lines and features of how someone drove the car, “explains Dan Greenwalt, creative director Forza In our latest war stories video, the franchise at Turn 10 Studios.

In fact, the technology originated at Microsoft Research’s outpost in Cambridge, England, where computer scientists started using neural nets to see if it was possible to identify a Formula 1 driver through corners. Get a computer to do it.
In the years since, Driveatars have evolved greatly – and not always that players appreciate. For the first few games, which were written for Xbox and Xbox 360 consoles, Driveatars resided entirely on the console’s local hard drive. As you lapped various tracks in the game, the driverate system was silently using that data to train the AI ​​you were running against in the background. But with the move to Xbox One, Greenvolt’s team decided to use the cloud.
“Suddenly going into Thunderhead, we were millions of laps within a week?” Says Greenwalt. “it was crazy.”

More data, more problems

Very quickly, in Turn 10 the team started learning a lot about how different players drove their cars. Tracks are not uniform – they vary in height, width, corner profile, and more – and this is complicated by the way neural nets try to normalize a racing line. Things get even more complicated when you’re racing against other cars; Human players are more aggressive with AI drivers than other humans, and by Forza motorsport 6The games began to include a way of dealing with single-player games to prevent some of AI’s aggression, which AI tried to ruin humans on every occasion.

In this video recorded remotely due to COVID-19 – Greenvolt goes into some recollection of how his team has handled these challenges for years and even discussed the terrible problem of “rubber banding” Is where AI cars slow down. Too little to allow a human player to catch him after a spin. Even if you are not a fan of racing games, we think you will like it very much.

Image listed by Ars Technica

Leave a Reply