'ATARI'

Hi there!

You've definitely stumbled upon the news shouting that Artificial Intelligence has finally beaten humans in this or that, or at least is close to us. Kasparov and Deep Blue, Atari 2600, Go, Doom, what is next? Now when we have plenty of environments like OpenAI Universe, Microsoft Minecraft Malmo, DeepMind SCII who knows what will happen tomorrow?

Doesn't it hit your pride? If yes, dive into this website and give AI a fight in Atari 2600 games: Q*bert, Video Pinball, Ms.Pacman, Space Invaders, and Montezuma's revenge. The AI best scores were taken from this Nature paper by Mnih et al.

More seriously, my name is Vitaly Kurin and I want to collect a dataset of human Atari games replays and use it for my master thesis partly connected to reinforcement learning research. As soon as the dataset is cleaned and polished, I will make it open, so that everybody will be able to use it for own research. So, by playing you accelerate the progress of human civilisation ;)

My advisors for the thesis are Sebastian Nowozin from Microsoft Research in Cambridge, UK, Bastian Leibe and Lucas Beyer from Computer Vision group at RWTH Aachen University.

I would also like to thank the creator of Javatari.js Paulo Augusto Peccin for the open source js Atari emulator and his help in answering many of my questions.

The first version of the dataset is finally released, you can find it here.

Disclaimer. Java, Atari and Video Computer System are trademarks of their respective owners. This project is completely non-commercial and created only for the benefit of the research community.

GO!