Earon Mask's Artificial Intelligence Research Institute OpenAI Releases AI Learning Platform "Universe"

Entrepreneurs and investors representing the United States gathered and founded as a non-profit research institute to open-source artificial intelligence (AI)OpenAI"is. Founded to make people around the world to benefit from AI, Mr. Ellon Mask who handles Space X and Tesla Motors and Mr. Sam Altman of Y Combinator president serve as chairman Although there is a software platform for measuring and learning AI's intelligence on December 6, 2016 in local time "Universe"Has been released.

OpenAI Universe
https://universe.openai.com/


Universe
https://openai.com/blog/universe/

Universe of AI learning platform developed by OpenAI is an existing libraryTensorFlowYaTheanoIt is a platform that can be used as it is.


In order to learn to AI it is necessary to have various tasks to be experienced, but Universe is a platform that AI agent "runs various tasks using a computer like a human being".


By having the AI ​​agent recognize the image displayed on the screen and using the keyboard and mouse on the virtual, we make human task to do various tasks to use the computer.


There are thousands of Universe's games and browser tasks to use for AI measurement and learning. Among games that AI uses for learningSlither.ioFromGTA VVarious kinds of things are available, and these are used as a type of machine learningReinforcement learningIt seems that we will advance learning of AI.


OpenAI provides an environment for training AI in April 2016 "GymHowever, since the newly released Universe can also change the program to Gym's environment, AI researchers who have used Gym so far can also move the environment. UniverseVNCIt automatically runs on Remote Desktop using, and there is no need for special access to the inside of the program, source code bot API,

Universe aims to provide a platform that can quickly learn tasks under difficult circumstances and unknown tasks, which is an important step to create AI with general intelligence. In addition, various companies such as EA, Microsoft Studios, Valve, Wolfram and others cooperate to construct the game library provided by Universe.


A movie shows that you are actually using the Universe to play the game on the AI ​​agent, and you can play the movie by clicking the following thumbnail.

In the Universe libraryAtari 2600The game title is recorded and it is possible to let you learn the AI ​​agent using this.


Furthermore, let's play flash game with more than 30,000 titles like this. Fresh games are higher in image quality than Atari 2600 games, but as games, there are still quite a few simple things.


It is like this when playing a racing game in a flash game.


In the case of games, AI agents will learn how to play better so that they can produce better scores. And, since there are many times numerical values ​​displayed on the game screen, we will learn so that better scores can be obtained using them. For example, in the screenshot below, the screen of the race game is displayed on the left side, and the variation of the score scouted by the AI ​​agent is indicated by the line graph on the right. You can see that the score simply uses the movement speed (MPH) displayed at the lower right of the game screen.


In addition to games, you can check how you are challenging a variety of tasks such as drag & drop which is frequently done on a web browser, entering characters into a text box, moving a slide bar and so on. If you click on it, you can see the actual motion in the movie.

The following movies are created for benchmark "World of BitsA movie that challenges tasks on the browser under the environment called "


The following movie is a movie that causes the AI ​​agent to run on a website that exists on the actual Internet and causes the task to be performed.


Although OpenAI is "a nonprofit research institute for making AI open source", I am publishing the source code of Universe on GitHub according to its policy.

GitHub - openai / universe: Universe: a software platform for measuring and training an AI's general intelligence across the world's supply of games, websites and other applications.