Get the latest tech news
The History of Machine Learning in Trackmania
Trackmania: the world's most competitive racing game. In this post, we begin our journey towards training a neural network to superhuman performance, by chronicling the history of previous attempts.
With a dataset of 100,000 images - equivalent to roughly a few hours’ worth of driving at ten frames per second - and with no concept of momentum or even any ability to use the brakes, it’s not surprising that the car eventually flipped out of the track and crashed during the test run. Three collector agents drive tracks, using a neural network with the game state (both the current screenshot and data such as position and velocity) as the input, and produces an output estimating the total reward the car is expected to experience throughout the rest of the run for each possible keystroke. The earliest, in 2019, was a French programmer named Bluemax666, who was inspired by a 2017 project in Python to play the game Grand Theft Auto V. He started the way many did: a simple program that could only accelerate and not brake, using a black-and-white bitmap of a small section of the screen, and converting manual keystrokes to virtual joystick inputs.
Or read this on Hacker News
