When will the day the artificial intelligence "DQN" that learns the game by himself learns more than humans threaten humans?

ByPhOtOnQuAnTiQuE

Reproduce the mechanism that the human brain learnsDeep learningThe advancement of research on artificial intelligence (AI) by computer has made a dramatic evolution.Google acquired 50 billion yenA British venture company said to have done "Deep Mind"Agent program developed by"DQN(Deep Q-Network) "once you start the program and learn how to hit the high score by self-practicing the game such as" Space invaders "and" block break "without permission by humans It is considered to be a breakthrough that further accelerates AI development in the future.

Google Deep Mind
http://deepmind.com/

Human-level control through deep reinforcement learning: Nature: Nature Publishing Group
http://www.nature.com/nature/journal/v518/n7540/full/nature14236.html

A movie played by Space Invaders by DQN developed by DeepMind is here. Although it may show slightly different movements from human playing,TorchkaYou can see how you steadily clear the stage by shooting down UFOs crossing the invaders and the upper part while protecting themselves.


DeepMind gave DQN 49 games and learned how to play, and eventually he succeeded in finding a higher score than human with 29 different games. In the following movie, learning the DQN which did not even know the rules of the game at the beginning, 100 times, 200 times while playing repeatedly "How to move to get points" You can see how it is.

Mnih Supplementary Data video 2 R3 - YouTube


By the time when the number of times of play reached 600 times, you can drill a hole in the wall aiming at the end of the block, send a pon (ball) to the upper space, and acquire a technique (tunnel) to get a large score . Finally DQN seems to have succeeded in getting a high score of 13 times that of humans.

In Deep Mind, DQN learning programReinforcement learning(RL: Reinforcement Learning) is adopted to improve the efficiency of learning. RL is to learn a strategy in which an agent in a certain environment obtains the most reward through a series of actions, which is exactly what humans try all kinds of strategies to target high scores in games It can be said the same way.

In this way, the development of artificial intelligence by computers continues to evolve at an accelerating rate, and in the future it can be said that it seems that it is no wonder that the time when the human brain is perfectly reproduced comes so. Ellon Mask who created Tesla and space X "Care must be taken in evolution of artificial intelligence"As I told you, in the future it may be necessary to think about the risk that computers will gain the ability to exceed human beings.

Meanwhile, Demis Hasabis, founder of DeepMind, says "It still needs more than 10 years to reach the threat." Although it is still ahead of the story, it may be only now that you can laugh with calling DQN "Docon".

Inside DeepMind - YouTube