Google DeepMind and Blizzard turn StarCraft 2 into AI learning environment

Google’s artificial intelligence research and development division has announced a partnership with Blizzard. We are talking about the collaboration of DeepMind and Blizzard on a project that allows you to turn StarCraft 2 into an environment for studying the weak form of AI. The plans of the company, including - the creation of a system that could defeat the champions in StarCraft.

DeepMind previously developed the AlphaGo system, which was able to defeat the world go champion several times . A few months ago, AlphaGo won the battle with Lee Sedoll. Jeff Dean, a spokesman for the Google research team, even before winning AlphaGo, announced his intention to engage in e-sports.

Speaking about the development of AI for StarCraft, the authors of this project mean a system that will receive the same initial data as a person. That is, this system will not know the location of enemy units and map elements until they are opened. The conditions for a person and a computer will be equal. Now the situation is somewhat different - you can fight with a computer, but he knows everything about the map, so you can’t call a war against such an enemy an honest battle.

Here is a demo of two StarCraft 2 champions:


In practice, this means that Google AI will need to predict what the enemy is doing at any given time in order to adequately respond to the actions of the “enemy”. In addition, it will be necessary to quickly respond to what has gone beyond the plan. All this will allow testing the next level of capabilities of artificial intelligence. to “In the end, we want to apply these abilities to solve global problems,” said Demis Hassabis, founder of Deepmind (it was bought by Google in 2014, and AI is being developed based on the achievements of the acquired company).

Google chose StarCraft because for 20 years this game has been the benchmark among esports players. Professional fights for StarCraft have been held since the late 90s of the last century, in fact, since the appearance of the game on the market. And now prize pools, the level of e-sportsmen and StarCraft championships are only growing.

DeepMind has long been engaged in research in the field of artificial intelligence. Developers are trying to solve complex problems in this area that were previously considered unsolvable. As an example, you can bring all the same game go. Experts claimed that computers could begin to play at a professional level in Go no earlier than in 10-20 years. And DeepMind managed to develop a system that emerges from the competition with the best of the best today. DeepMind now works not only with StarCraft, but also with other games, including two-dimensional old games with Atari or Torcs, an open-source car racing simulator.



Starcraft- One of the first real-time strategies. She appeared on the market in the 98th year. Here, each side differs not only in appearance, but also in the characteristics of its representatives (there are three races in total - terrans, protoss and zerg), as well as a common battle strategy. Protoss have powerful, but expensive warriors and equipment, and zerg outperform the enemy due to their numbers and speed. All things being equal, no race has a significant advantage over another. These conditions include the number and location of resources in close proximity to the first base, the defensive conditions of the area, and the location in relation to other players.

In order to be able to win at StarCraft, the player needs to show the efficient use of the resources of his race. He must also be able to plan, drawing up short-term and long-term plans, which, moreover, must be flexible enough so that they can be changed when new factors appear. AI will have to learn how to control the mouse, camera and be able to work with game resources.

The joint developments of DeepMind and Blizzard will become available to researchers in the field of games and artificial intelligence as early as next year. Partners create an API that allows the use of bots, the principle of which is similar to those programs that some users use now. In this case, opponents will be able to work with the game environment in the form of pixels, from there with the most simplified game cards in RGB format. The following is an example of what the game layer might look like when using the API.



“Computers can run very fast, but this doesn’t mean intelligence at all,” said Oriol Vinyals, a researcher at DeepMind. He also claims that so far, the artificial intelligence developed in the company is far from the level of the game that modern champions show. But we must not forget that this company created the AlphaGo system relatively quickly, and after its presentation, many experts claimed that it could not play on equal terms with Lee Sedol. Indeed, playing on equal terms did not work out - the computer won 4 times out of 5, showing a higher class of the game.