DeepMind AI agents conquer human professionals in Starcraft II



The artificial intelligence agents developed by the Google affiliate DeepMind have hit the human professionals in Starcraft II – A novelty in the world of artificial intelligence. In a series of games aired on YouTube and Twitch, AI players defeated humans 10 games in a row. In the final match, professional player Grzegorz "MaNa" Komincz was able to wrest a single victory for humanity.

"The history of AI has been marked by a series of significant reference victories in different games," said David Silver, co-leader of DeepMind's research, after the games. "And I hope, although there is clearly work to be done, that people in the future can look back [today] and maybe consider this as another step forward for what artificial intelligence systems can do. "

Defeating humans in video games may seem like a secondary spectacle in the development of artificial intelligence, but it is an important research challenge. Games like Starcraft II They are harder to play for computers than table games like chess or Go. In video games, artificial intelligence agents can not see the movement of each piece to calculate their next move, and they have to react in real time.


A screenshot of the games in December, which shows AlphaStar facing TLO.
Image: DeepMind

These factors did not seem to be an impediment to the DeepMind AI system, dubbed AlphaStar. First, he beat the professional player Dario "TLO" Wünsch, before moving on to face MaNa. The games were originally played in December last year at DeepMind's headquarters in London, but today a live match was broadcast live against MaNa, which gave humans their only victory.

Professional Star boat commentators described the AlphaStar game as "phenomenal" and "superhuman". Starcraft IIPlayers start on different sides of the same map before building a base, training an army and invading enemy territory. AlphaStar was particularly good at what is called "micro", short for micromanagement, referring to the ability to control troops quickly and decisively on the battlefield.

Despite the fact that human players sometimes managed to train more powerful units, AlphaZero was able to overcome them in close maneuvers. In one game, AlphaStar invaded MaNa with a fast-moving unit called Stalker. Commentator Kevin "RotterdaM" van der Kooi described it as "control of the phenomenal unit, it's just not something we see very often". MaNa observed after the game: "If I play with any human player, they will not be cheating on their Stalkers so well."

This reflects the behavior we have seen in other high-level AI games. When OpenAI agents played human professionals in endow 2 last year, Ultimately, they were defeated. But experts said the agents again played with "clarity and precision" that was "hypnotic." Making quick decisions without any mistakes is, as expected, the terrain of a machine.

Experts have already begun to badyze the games and discuss whether AlphaStar has any unfair advantage. The AI ​​agent was limping in some way. For example, I was restricted from making more clicks per minute than a human. But unlike human players, he was able to see the entire map at once, instead of manually navigating it.

The DeepMind researchers said that this offered no real advantage, since the agent only focuses on one part of the map at a time. But, as shown in the games, this did not stop AlphaStar from expertly controlling units in three different areas simultaneously, something that commentators said would be impossible for humans. In particular, when MaNa beat AlphaStar in the live game, the AI ​​was playing with a restricted camera view.

Another potential pain point included the fact that human players, while professionals, were not world champions. TLO, in particular, also had to play with one of Starcraft II & # 39; s Three races I was not familiar with.


A graphic representation of the AlphaStar processing. The system sees a complete map from top to bottom and predicts which behavior will lead to victory.
Image: DeepMind

Leaving aside the discussion, experts say that the parties were an important step forward. Dave Churchill, an AI researcher who has been involved in the Star boat AI scene, he counted The edge"I think the strength of the agent is a significant achievement, and it came at least a year ahead of the most optimistic conjectures I've heard among AI researchers."

However, Churchill added that since DeepMind had not yet published any research work, it was difficult to say whether or not it showed any technological advance. "I have not read the blog article yet or I have not had access to any documents or technical details to make that call," Churchill said.

Mark Riedl, badociate professor of artificial intelligence at Georgia Tech, said he was less surprised by the results and that this victory had only been "a matter of time." Riedl said he did not think the games showed that Starcraft II He had been definitely beaten. "In the last live game, restricting AlphaStar to the window eliminated part of its artificial advantage," Riedl said. "But the biggest problem we've seen … is that politics learned [by the AI] it is fragile, and when a human can push the AI ​​out of its comfort zone, the AI ​​collapses. "

Ultimately, the ultimate goal of a job like this is not to defeat humans in video games, but to refine AI training methods, especially to create systems that can operate in complex virtual environments such as Star boat.

To train AlphaStar, the DeepMind researchers used a method known as reinforcement learning. Agents play the game essentially by trial and error while trying to achieve certain goals such as winning or simply staying alive. First they learn by copying human players and then they play each other in a competition similar to a coliseum. The strongest agents survive, and the weakest are discarded. DeepMind estimated that its AlphaStar agents accumulated about 200 years of game time in this way, played at an accelerated pace.

DeepMind was clear about its goal in carrying out this work. "The first and most important of the mission in DeepMind is to build an artificial general intelligence," said Oriol Vinyals, co-leader of the AlphaStar project, referring to the search for an artificial intelligence agent that can perform any mental task that a human being can "To do so, it is important to make a comparative evaluation of the performance of our agents in a wide variety of tasks."


Source link

Check Also

OnePlus latest teaser suggests ‘Nord’ branding for your next budget phone

After officially confirming plans to launch a new affordable smartphone series, OnePlus has now hinted …