DeepMind AI triumphs. Artificial intelligence beat StarCraft II players

Fielding AI against professional players is not a new practice. Some of you probably remember the OpenAI that Danil “Dendi” Ishutin had to deal with, a great DOTA 2 player, currently outside of the main Natus Vincere team? If not, I encourage you to return to this material, the link is below in the text. This time it fell on two professionals from the StarCraft II scene. They are the ones who decided to defeat an artificial creation of Google with a rather peculiar name of DeepMind AI. Unfortunately for them, the mentioned AI turned out to be better.

The whole action may be even more important for you, mainly because one of the two professionals was our compatriot. Grzegorz “MaNa” Komincz represents the colors of Team Liquid and it was he who took part in the fight against DeepMind AI. As one of the top players, commanding the Protoss army, he decided to take on the AI ​​who had previously learned his licks. With the right amount of data, the AI ​​knew exactly what it needed to do for MaNa to fail. Grzegorz and Dario “TLO” Wünsch played five games and it was quite an experience. If there was a human on the other side, he could count on similar techniques, specific movements and a certain repetition in actions. Unfortunately for them, each game required a whole new approach. DeepMind was not schematic and changed style often so the rival couldn’t predict next moves.

The training plan for this artificial intelligence can translate to 200 years, where all the AI ​​did was play StarCraft II. No one would be able to do something like that, and for the “machine” it’s just a task. It would seem that such an inhuman being will surpass his flesh-and-blood rivals in every way. This is the number of actions per minute, as well as the speed of reaction to given events. This was not the case with DeepMind AI. The creators explained that the APM (action per minute) of their works averages 280, and the response time is around 350 milliseconds. These results are only slightly above what the “average” professional player can do. The path to success was completely different and you don’t need to be a faster clicker to defeat your human opponent.

If you have been in contact with this game, you know well what you need to keep in mind to win the game. In addition to the basic building, we have a plethora of different units with a specific task: harvesting minerals, attacking the enemy from the ground or the air, protecting others, healing, etc. There is also building new fortifications, managing the army, upgrading characters, vehicles and all buildings. We must monitor both what is happening on the ground and in the air. Let’s not forget about the topography of the area, the location of resources on the map and, of course, the enemy himself. There’s a lot of thought here, and I admire how the professionals handle it all. A hesitation, a stupid gesture, too much haste or a badly executed attack will quickly make the rival an advantage. The game is not the easiest, as the threshold for entering online mode can be quite high. Professionals have mastered a lot of things, and that does not guarantee that they will be able to win the tournament. The most impressive is the artificial intelligence that has been able to cope with what StarCraft II offers.

Watch also: Artificial intelligence beats competitor NaVi

Instead of focusing on even faster response and higher APM, Google’s work focused on something else. First of all, it reads the whole interface completely differently from a human who has to choose where to place the camera at any given time. Although AI behavior is based on human practice anyway, it works a bit differently. Second, it handles everything much better. We are talking here about units, resources and movements made on the battlefield.

Because of this, it may react a bit slower than professionals, but still, DeepMind AI actions are more efficient. Finally, it should also be kept in mind that the “machine” has studied the performance of these players. She learned their licks, got to know the good and bad sides of the strategies used. With so much data, she was able to win with them. The whole process was very interesting for the participants of this contest. Grzegorz was impressed with what AI can do.

During the matches, she was able to use completely different strategies while remaining completely “human”, which the Pole did not expect. He realized that his game forces his opponent to make mistakes and predict their reactions, which puts everything in a new light. MaNa won with the machine once, but it was a new version, with a “different interface”, where everything is not yet perfect. version with a restricted field of vision. Before that, the machine was aware of everything that was happening on the map. When DeepMind had to pan the camera like a human, things changed.

I’m very curious if training against such an AI would help me compete with other professionals? Can you learn machine techniques that others don’t expect? In theory, it’s possible, but how will it work in practice? However, that doesn’t change the fact that we’ve been living in the era where artificial intelligence is able to manage 200 units on a large map and eliminate its human rival. It’s not Skynet, but just another step towards “real machines”. Matches in StarCraft II are mostly 1v1, but you can also play 4v4. Of course, DeepMind AI isn’t ready for something like that yet, but it will be in the future.

Leave a Comment