DeepMind’s AlphaStar AI Achieves StarCraft II Grandmaster Leaving Pathetic Humans In Its Wake
Some professional gamers have conflicting emotions about AlphaStar claiming Grandmaster status in the game. Despite any mixed feelings people have about the AI's new status, DeepMind says that AlphaStar's development will help it to develop AI tools that could benefit humanity. DeepMind notes that one of the things that has it excited about Starcraft II is that the game poses challenges that are seen in the real world.
DeepMind says that Starcraft is a "benchmark domain" to understand the science of AI and advance the quest to build better systems. Examples of technology that might come from its work in AI science include robots, self-driving cars, and virtual assistants. DeepMind notes that all of those systems need to make decisions based on "imperfectly observed information."
Part of the challenge of real-time games like Starcraft II is that players aren't taking turns making their moves. Everything is real-time, meaning an attack could come at any moment. At any given time in the game, there are up to 100 trillion trillion possible moves. DeepMind created three neural networks, one for each alien race in the game.
The AI was trained using a database of old games Blizzard offers. The entire AI training lasted 44 days, but due to the high speed, it represents about 200 years of human gameplay. The result is that the neural networks have Grandmaster status for each of the three races. There are still between 50 and 100 human players who can beat AlphaStar. DeepMind announced its AI would take on Starcraft II in 2016. The AI began to secretly fight human players in public ranked matches in July 2019.