Home > News content

Facebook secretly took part in a AI StarCraft game, and he lost the game

via:博客园     time:2017/10/11 17:06:49     readed:876

After Google's Alphago shine, the search giant aimed the next goal at a higher level of "StarCraft" game. But the social giant Facebook is not far behind, shortly before itsecretlySent a robot to participate in an interstellar battle between AI. Despite the fact that most of the rules-based robots are made with lovers, Facebook makes the robot's own strategy for the future.


In the galaxies of the galaxy, Facebook's group of Zerglings is hovering outside the enemy base. In the head of the lack of tendons of the commander to open the door, the social network of aliens flocked in a battle for artificial intelligence research positions in the battle stationed in the inside of the troops to kill.

This bloody event is part of an annual StarCraft competition by AI software robot, this year's game just ended on Sunday. Facebook quietly sent a robot named CherryPi, which was designed by eight people with its AI research lab.

This secret space station shows that Facebook and Google and so on the first to achieve the next milestone in AI's milestone in the competition is serious. Google's DeepMind AI research department posted a news headline last year because Alphago defeated the Go World Championship. In August of this year, DeepMind announced the latest version of StarCraft II as its next goal.

Like most of the AI ​​research in this field, Facebook's participation in this game is using the older version of StarCraft, this version is difficult to think of software and players are treated equally. Facebook's AI research department is led by New York University professor Yann LeCun, with more than 80 researchers, and has published a number of research papers, but has not yet achieved any shocking achievements like Google's Go. About StarCraft Facebook has published three research papers, but has not yet announced any special action to conquer this game.

The final result announced on Sunday showed that Facebook still had a way to go: CherryPi ranked 6th in a total of 28 players and the top three were written by independent programming enthusiasts.

Facebook research scientist Gabriel Synnaeve said CherryPi as the future of StarCraft research's "baseline". He said: "We want to see it with other existing robot comparison, in particular, want to test is not there need to amend the flaws. CherryPi participated in a marathon race, which is part of the AIIDE study of AI's academic conference in the entertainment industry. Facebook also sponsored this year's competition, thousands of robots against the game hardware used by Facebook money.

Since the 1950s, games like chess, checkers, chess and chess have been a test stand for new ideas for artificial intelligence. Since this day, as the company increasingly use AI as a means of grinding their own products and services, AI also serve the serious business purposes. Facebook, Google and other technology companies use AI to improve targeted advertising and personalized systems, and to virtual assistants and enhance the reality and other new products to provide support.

The temptation of StarCraft to AI researchers is not just the fun of manipulating weapons. Although the game seems to be more approachable than chess or chess, but the complexity is much higher than the latter, because the player's equipment and actions are not limited to a fixed board and in the other side of the full view of the following. The effective number of chessboards is 1 followed by 170. Researchers estimate that in order to traverse the complexity of StarCraft, you will need to add 100 more in that number.

This year's winner called ZZZKBot, the developer of Perth, Australia's software developer Chris Coxe, who had worked in Nasdaq before. The robot was developed by himself alone, and later in order to set aside more time to do this he even stopped working for some time. One day before the final result was released, Coxe had ridiculed his own work. He said: "This thing is only used to verify the concept." Source code is not good. & rdquo;

Just like all StarCraft robots so far, in the game skills in general human players in front of ZZZKBot also not long for too long. The planning and memory needed to anticipate and respond to the mobilization of alien forces is far beyond the current capacity of the software.

With the two giants competing in online advertising and AI competing at the same time, enthusiasts have made the best StarCraft robot days seem to have been numbered. David Churchill, the organizer of the AIIDE contest, and the Newfoundland Memorial University professor, predicted that the situation of StarCraft robots will change dramatically over the next few years.

Facebook and Google said they developed the StarCraft robot approach with most of the individual programmers are not the same. Most of the leading robots are based on the rules and strategies developed by the creator. Coxes said one of the best features of his robot is a simple learning function, and the robot will try pre-programmed strategies for other Battle robots and then record which strategies are effective to prepare for the next game. The two technology giants intend to rely more on machine learning, plans to let the robot rely on checking the past game data cache or repeat the experiment from scratch to develop their own strategy. Facebook did not develop CherryPi according to his own ideas. And machine learning is Google's Alphago invincible core.

Although Facebook's robot did not win the Star Wars, but the second-ranked robot PurpleWave developer Dan Gant saw CherryPi future trend. The choice of most robots is based on the relative number of enemies either a positive attack or a retreat. But Gant said CherryPi seems to know when he can attack the base quickly enough to see the video released before the results are announced.

Nevertheless, do not expect independent robot developers to disappear overnight without a trace, and expect StarCraft will soon be conquered. Churchill said: "The problem is still very difficult. I expect a few years of lovers to do the main rule-based robots that will still perform well. "He guessed that any robot would have to beat the expert level players would still take five years [mdash; & mdash; but admitted maybe it would be faster.

New York software developer Gant has stopped working this hand for a few months to do PurpleWave. He said that the addition of technology giants has added an influence that represents the pursuit of unique learning opportunities. He said: "Whether you are Facebook or DeepMind or just learning to learn the child, we are in the same arena fair sports." Can limit your own only their own efforts and can teach their own things. & rdquo;

To make a StarCraft super players can bring to the technical giant is not just a sense of satisfaction. Google said DeepMind's machine learning has helped it cut the cooling costs of its own data center. Microsoft this year, a study of machine learning papers that predict the user will click on the accuracy rate even if only raised 0.1 percentage points can bring hundreds of millions of dollars in new income. Have the ability to lead the Zerg team to lead the crash to defeat any human robot may soon be able to make a lot of money.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments