"Gym Retro" which allows you to learn the retrogame capture by AI and create SuperPlay

Established as a nonprofit research institute on AIOpenAILet AI learn the capture of retro gamesReinforcement learning (RL)Platform for "Gym Retro"Has been released. By using Gym Retro, we can use it for academic research of RL, and we can make effective game strategy method AI.

Gym Retro
https://blog.openai.com/gym-retro/

According to OpenAI, Gym Retro is developing to study the RL algorithm, and it is possible to learn not only one game strategy but also multiple games by one AI.

Also, the games that can be used for learningMega drive,Sega · Master System,NES,Super Nintendo EntertainmentYaGame BoyCompatible with a wide range of platforms such as. theseGame emulatorIf you prepare a ROM image to work with, you can let Gym Retro learn the game.

By using the following learning tools, Gym Retro lets you learn the "remaining machine" and "score" of the game ... ...


It is possible to let AI play various games.


However, if you let AI learn with the aim of only "raising the score" in action games etc., after defeating the enemy as follows ...


Move the stage twice and move to the original stage.



And it sometimes falls into an infinite loop that defeats the same enemy and moves again. Therefore, OpenAI says that learning needs to be devised and done.


If you can let Gym Retro successfully learn it, it is not a dream to AI to create the optimal procedure that nobody ever noticed, and to clear no-damage. Especially in the case of scrolling games like Gradius, the difficulty level of learning is said to be relatively low.


Gym Retro is compatible with Windows, Linux, macOS, and to operatePython3.5 or 3.6 is required. How to install,Documentation of Gym RetroIt is explained in.