An impressive new victory for artificial intelligence not just over humans, but over the owners of professional skills on their field of play. AI "AlphaGo" first got acquainted with the team shooter Quake III and after a series of training was able to consistently defeat all human rivals, from ordinary to pros. But what is even more interesting, no one taught him the intricacies of the game, the AI itself figured out everything and even mastered the cooperative style at a level inaccessible to people.
The tests took place in the "Capture the Flag" game mode. It should be noted right away that the AI was put in the most disadvantageous position on the fly by connecting it to the game interface without interpreting the data. That is, the AI saw pixels moving on the screen and knew only one criterion for success - placing a flag at a certain point in space. The concepts of "friend", "enemy", "base", "defense", "counterattack", even the presence of a weapon in the game and how to use it - all this AlphaGo deduced logically, observing the process. It took him 140, 000 sessions to get comfortable with the game at the level of a regular gamer through trial and error.
But then everything went much better - after 175, 000 trainings, the AI began to consistently beat professional players. After 400, 000 matches played, the best played pair of gamers, who studied AlphaGo's style for 12 hours, and actively communicated with each other during the competition, ended up losing to him with a score of 1: 4. Other teams lagged behind the AI by 16 flags, and in the tournament, almost forty people voted AlphaGo a "better teammate" than human players.
Note that the game involved a new version of AlphaGo, with a double level of "thinking": analytical, for developing a strategy, and practical, for making immediate decisions. He used the first when he came up with, and then rejected several tactics, and with the help of the second he created his own markers and detectors of important objects and events in the game. Trying to complicate the task, the researchers lowered the AI's firing accuracy from 80% to 50%, and the response time was stretched to 267 ms in order to neutralize the superiority in information processing speed over a living person. It didn't help - AlphaGo won anyway.
The most intriguing fact: the best combination in the game "capture the flag" was a duet of man and AI, and they did not contact each other and did not coordinate their actions. It turned out that it is easier for an AI to adjust its tactics to the actions of a leading human than to organize the work of two AI players. And even if the difference between the pairs was only 5%, it is obvious. Our world is approaching the era of creating ideal companion robots for humans.