The artificial intelligence learns to read videos of Atari

Over the past few months there have been great advances in the field of Artificial Intelligence (AI). But sometimes engineers can afford to teach trivial things to these platforms; how to play Atari.

This just happened with the friends of OpenAI, who presented the results of their latest project in the field. Where they taught an AI to play the clbadic Atari, Montezuma's Revenge.

To succeed in a reinforced learning project, an AI must do two things:

Find a sequence of actions that leads to a positive reward. This is the problem of exploration.
Remember the sequence of actions to take, and generalize it to related situations; but slightly different. This is the problem of learning.

This was the base points of the project of OpenAI. Where the best way to achieve their goal was to make video recordings.

Different members of the video team were recorded playing sessions a little over ten minutes of the game. So Artificial Intelligence badyzed the games and learned what to do to move forward in each scenario.

Thus, the human demonstration of the game managed to obtain 71,500 points at the end of its session. Artificial intelligence has imitated the road; but he did 74,500 points.

This may seem inconsequential. But that represents an important step forward. Especially that next month, this same platform will play Dota.