Training an Unbeatable AI in Trackmania
Training an Unbeatable AI in Trackmania by Yosh
About the Game
Trackmania is a popular racing video game series that dates back to 2003 with 15 separate titles published on multiple platforms up to this date. What sets this game apart is its unique feature that enables players to create and share their own tracks within the game, fostering a vibrant community of track designers.
Trackmania Nations Forever
Among these 15 games developed over the span of 20 years, the most cherished and widely played by the community is Trackmania Nations Forever which came out in the year 2008. Nations Forever is a free PC game that despite being an old game, is still played by many players trying to break records in both official and community-made tracks. The reason why people still play this game can be attributed to its crisp physics and its simple and fun gameplay.
One person managed to train an AI that can not only play this game but also beat skillful players while doing so.
AI vs Humans
The concept of training an AI to beat humans in games is not something new. From traditional games like "chess" and "go" to modern massive competitive video games like "Dota 2" and "Starcraft II", AI trained on a huge scale managed to beat the pros of these games.
OpenAI's Dota 2 bot beating a top player
So, what made Trackmania Nations Forever's AI improvement so interesting then? (4 million views on YouTube in 10 days from a channel with 80k subscribers indicates that it is.)
Aside from the beautiful presentation of the video, the technical reason would be that the AI is not easy to train because of the fact that this game is dependent on physics that are close to real life which, unlike the other aforementioned games, do not need to deal with. It should be also mentioned that this is a one-person project so it makes it more impressive.
Training Process
Although the video explains the concepts very well, I would like to sum it up with my opinion for the readers.
The creator of the video "Yosh" made 3 levels of tracks that have no vertical slopes. Yosh is a skilled and experienced Trackmania player who has been playing the game for 17 years. He has been trying to train a model like this for 3 years and he explains his process of doing that in 6 months.
The training method for the AI was reinforcement learning where simply put, the model is rewarded for doing something considered "good" and hence, reinforcing that behaviour. In this case, this good action is carrying over speed over sections of the track. He usually used trial and error for fine-tuning the model parameters. After all the efforts, he got himself an AI that he could not beat.
This of course is not the end of the AI's journey on Trackmania, as it is apparent that the game gets more complex as the tracks get crazier (and believe me they get far crazier).
Concluding
One thing that can be taken from this video is that to be able to be successful with such models, you need to understand the core mechanics and techniques in the game very well. Also to succeed, one must possess passion, determination, and patience.
Thanks for reading!
Oğuz Arslan
Comments
Post a Comment