Playing the game also lost! The new "Alpha" beats human players

Playing the game also lost! The new "Alpha" beats human players

tenco 2019-01-28

After the "man-machine war" of go successfully defeated the top human go master, Google's artificial intelligence technology company "deep thinking" artificial intelligence program "alpha star", and the human professionals in the classic real-time strategy computer game "starcraft 2" against, won a landslide victory.

Since December, the company's alphastar artificial intelligence program has competed 11 times against high-level human professionals, according to deepmind.

In this series of "man-machine war", the artificial intelligence program finally defeated the human player 10 to 1.

Only in the live competition on January 24th did the human players eked out a win because the alpha star's view was limited.

David silver, deep thinking's co-head of research and development, said after the competition: "while there is still a lot of work to be done, I hope that when people look back on today, they will see this result as a step forward in the capabilities of ai systems."

According to the report, unlike all the pieces on the go board, which are visible to both sides, there is a "fog of war" in such games, in which one side needs to guess and detect the other's actions, which belongs to the "imperfect information game", and requires artificial intelligence to respond in real time, which has higher requirements for artificial intelligence.

Deep thinking says that before alpha, no artificial intelligence system could match the skill of a human professional in starcraft ii.

"Alpha centauri" to the "I" is an important foundation of cuhk outscored depth is that it USES the neural network, the researchers through the way of supervision, learning and reinforcement learning, original data to train the direct use of the game, to imitate the thinking of learning, let the model quickly learn to high levels of human players in the game use strategy and operations.

In addition, games like starcraft ii have a concept called "hand speed," which is measured by APM (number of instructions per minute).

In the benchmark test, alpha star was able to execute about 280 instructions per minute, far less than a human professional, but with more precise movements that helped win the race.

What will winning go and computer games bring to the development of artificial intelligence?

"Deep thinking" of the research team believes that training "alpha centauri" advanced method and algorithm of the advanced architecture of the future will help researchers to accumulate more experience, and ultimately designed to deal with a lot of complicated problems in the real life of artificial intelligence programs, such as weather forecast, climate model calculation, and language understanding, etc.