Kamil Rocki was inspired by the 2016 paper from Google Deepmind researchers explaining how they used machine learning to develop a system that could play Breakout on the Atari 2600 with superhuman proficiency.
But Rocki wanted to develop a more general approach to machine-learning game-playing, one that could play more sophisticated games than Breakout, with less preprocessing and other time-consuming steps.
This led him on a quest to find a games platform with a lot of available games, that ran on simple enough processors that he could emulate a lot of them, very quickly, on modern computing hardware.
And so the Gameboy Supercomputer was born, running at over a billion FPS. He's successfully used it to develop machine learning systems that can beat Pac-Man and Mario Land.
Rocki finishes his writeup with a roadmap for further work from the Open AI community.
This is an excellent example of the way that flexibility in copyright -- including the freedom to rip old games and play them in emulators -- will be key to producing better machine learning systems.
In order to make learning more efficient, you could imagine trying to transfer some knowledge from simpler games. This is what remains unsolved right now and is a hot research topic. A recently published challenge by OpenAI tried to measure just that:
1. There is no obvious score
2. If no action is performed it takes 60 minutes to end the game (58 minutes left in the animation).
Could you try the exact same approach as in the Atari 2600 paper? I you think about it, how likely is it that you’ll get to the end by pressing keys at random?
A GAMEBOY supercomputer [Kamil Rocki/Toward Data Science]