Years ago, IBM defeated Garri Kasparov by playing chess; after Google did the same with Fan Hui and Lee Sedol something. Now, it's been the turn of eSports with one of its great games, the StarCraft II. The Artificial Intelligence (AI) tool developed by DeepMind, Google's subsidiary, has managed to defeat for the first time two champions of this video game.
The program used has been AlphaStar. With him, DeepMind has developed a deep network of neuronal learning directly trained through raw data from StarCraft II, as explained by the Google company in a Deepmind web post.
In a series of tests that took place on December 19, AlphaStar managed to beat the Polish player Grzegorz Manna Komincz and his partner, the German Dario TLO Wünsch, both members of the Team Liquid professional team. This game took place on a competitive map of the game and without rule restrictions.
DeepMind had achieved success with other titles such as Atari, Mario and Dota 2, and, since last summer, also Quake III in the mode capture the Flag. However, the difficulties of the Starcraft had been a test too hard. Until now.
The series of video games StarCraft belongs to Blizzard. It premiered in 1998 and is a real-time strategy game. DeepMind as a "great challenge" due to the complexity of the game, its mechanics and the breadth of the maps, which makes it difficult to train automatic systems to be competitive. In Starcraft, unlike what happens in chess or Go, the player does not have all the information before making the move, since the map is unveiled as the units move forward. Therefore, having intuition and imagination and cognitive skills are key to trying to guess what the opponent is doing. That is, everything that artificial intelligence lacked.
There are several different ways to play the game. In eSports, however, the most common is a one-on-one to five-game tournament. To begin with, a player must choose to play one of the three extraterrestrial "races": Zerg, Protoss (the race in which Alphastar has been trained) or Terran. All of them have distinctive characteristics and abilities. Professionals tend to specialize in a single race. Each player begins with a series of work units, which gather basic resources to build more units and structures and create new technologies.
These, in turn, allow a player to harvest other resources, build more sophisticated bases and structures, and develop new capabilities that can be used to outwit the opponent. To win, a player must carefully balance the long-term global management of their economy, known as in real life as macro strategies, along with the low-level control of their individual short-term units, known as micro.
Deepmind explained that StarCraft is a game in "there is not a single best strategy". In addition, also unlike Go or chess, players do not alternate in the game, but make decisions in real time continuously.
To train your AI, DeepMind has used raw data from the StarCraft II interface through two techniques known as supervised learning Y reinforced learning. The neural network converts the game units and uses an LSTM memory core that provides support for long-term learning, reports Europa Press.
Google algorithm a multiple learning algorithm that has been used initially to train AlphaStar's neural network through supervised learning, with which it learns from human players from other Blizzard video games and from its economy system macro and resources micro. With these techniques, he defeated the 'elite' difficulty of the studio games 95% of the time.
Subsequently, the researchers subjected AlphaStar to a reinforced learning process, for which a continuous StarCraft II league was created with real players competing, which created a global map with the strategies chosen by the people.
Subsequently, AlphaStar analyzed the success rate of each strategy and its possible counterattack tactics. With the StarCraft league, AlphaStar has accumulated an experience of more than 200 years of real game, added over 14 days.
The Google system managed to beat the professionals 'MaNa' and 'TLO', something that happens for the first time with eSports players according to DeepMind, and also with a result of 5-0. For this, it took advantage of a greater average of actions per minute, of tens of thousands against hundreds, and overcame to limitations of the algorithm like a delay of 350 milliseconds between observation and action.