The boundary line between a game in which Google's artificial intelligence "DQN" can play better than a human and a game that can not be done

Google's exclusive game artificial intelligence "DQN(Deep Q-network) "learning while artificial intelligence himself plays the game, it is possible to improve skill improvement more than human beings. As a result of such DQN playing various games, it became clear that there are games that can be played more successfully than humans and games that are not so.

Google DeepMind Blog Deep Reinforcement Learning
https://deepmind.com/blog

Human beings can solve a wide range of tasks ranging from simple tasks such as body motion control to high level cognitive tasks. Google's artificial intelligence development company "DeepMind" aims to develop artificial intelligence with the same level of performance and universality as humans. Because "DQN" of game specialized artificial intelligence developed by DeepMind aims at goal by repeating trial and error many times during game play, its learning form isReinforcement learningIt is called reinforcement learning. In addition, DQN self-learns and accumulates knowledge.

In DeepMind, DQN let me play 50 kinds of Atari 2600 games with no prior knowledge, and as a result, DQN showed more play than human's skill in 29 games about half. The game with the highest score "Video Pinball"It is 2539% higher compared with human beings. Even compared to the play result of another linear learning model shown in a gray graph, it indicates the high performance of DQN. On the other hand,"Montezuma's Revenge" game where the main character dies quickly.It fought very hard, and the play result of DQN is "0%".


Incidentally,DQN source codeWhenAtari 2600 emulatorBecause it is open to the public for free, anyone can reproduce DeepMind's experiment.

The state that DQN plays an Atari 2600 block breaking game "Breakout" is from the following movie.

DQN Breakout - YouTube


Also, you can see how Space Invaders play.

DQN SPACE INVADERS - YouTube


DeepMind improved DQN's algorithm, stabilized learning mechanics, handled prior gameplay experience preferentially, and standardized, gathered and remeasured the output. With this improvement, it seems that Atari 2600 game score has tripled or more. Also, since the game of Atari 2600 is limited to 2D space, he seems to have independently developed a game "Labyrinth" aiming for a goal in a 3D maze and introduced it for learning DQN.


Actually you can see how DQN is playing Labyrinth in the following movie.

Asynchronous Methods for Deep Reinforcement Learning: Labyrinth - YouTube


In Deep Mind, we are pursuing development so that artificial intelligence such as DQN can perform various tasks, continue to develop, continue to develop artificial intelligence ability, seeking ways to benefit society, such as health care applications ... apparently ...