Following on their victory in StarCraft II, Google’s DeepMind researchers lay out a plan to find diversity in players of games to get around the problems of hidden information in game theory. Their work points to a kind of “open-ended learning” that surpasses simple test-taking found in most machine learning.
The miracle of AI in the realm of playing games is encapsulated in the AlphaGo Zero program, which in 2017 was able to beat all human players of the ancient strategy game strictly by what’s called “self-play,” exploring different possible moves without any human tips. A revised version, AlphaZero, gained such general knowledge that it can now excel at not only Go but also chess and the game Shogi, Japan’s version of chess.
Hence, neural nets can generalize across many games just by self-play.
But not all games, it turns out. There are some games that don’t lend themselves to the AlphaZero approach, games known technically as being “intransitive.”
For these games, Google’s DeepMind AI researchers have figured out a new trick, a way of constructing a kind of super athlete by seeking diversity of moves and styles of play. The primary example, AlphaStar, recently won a match against the best human player of the strategy video game StarCraft II, as ZDNet chronicled last week.
We offer FLIE Blog updates to offer subscribers a weekly digest of posts made on the FLIE site. To subscribe to these email updates, simply subscribe below.
