Google's DeepMind AI Plans To Take On StarCraft II (venturebeat.com)

An anonymous reader quotes a report from VentureBeat: Google and Blizzard are opening up StarCraft II to anyone who wants to teach artificial intelligence systems how to conduct warfare. Researchers can now use Google's DeepMind A.I. to test various theories for ways that machines can learn to make sense of complicated systems, in this case Blizzard's beloved real-time strategy game. In StarCraft II, players fight against one another by gathering resources to pay for defensive and offensive units. It has a healthy competitive community that is known for having a ludicrously high skill level. But considering that DeepMind A.I. has previously conquered complicated turn-based games like chess and go, a real-time strategy game makes sense as the next frontier. The companies announced the collaboration today at the BlizzCon fan event in Anaheim, California, and Google's DeepMind A.I. division posted a blog about the partnership and why StarCraft II is so ideal for machine-learning research. If you're wondering how much humans will have to teach A.I. about how to play and win at StarCraft, the answer is very little. DeepMind learned to beat the best go players in the world by teaching itself through trial and error. All the researchers had to do was explain how to determine success, and the A.I. can then begin playing games against itself on a loop while always reinforcing any strategies that lead to more success. For StarCraft, that will likely mean asking the A.I. to prioritize how long it survives and/or how much damage it does to the enemy's primary base. Or, maybe, researchers will find that defining success in a more abstract way will lead to better results, discovering the answers to all of this is the entire point of Google and Blizzard teaming up.

Comments

One huge difference

[Solandri]

Games like Chess, Go, Tic-Tac-Toe always let both players see the complete world state. Armed with that knowledge, it's easy to be systematic and deterministic.

Games like Poker and Starcraft hide part of the world state from each player, forcing them to guess at the parts they can't see. That opens up the possibility of one player bluffing - leading the opponent down the wrong decision tree because he's fooled into thinking the part of the world state he can't see is different from what it really is. I don't think this is something an AI can "solve". Certainly one could optimize it, so that it becomes damn good at guessing when a certain player is bluffing or not. But put it up against a different player and all that "learned" experience becomes useless, or even counter-productive. Or even pit it against the same player who's aware he's playing against the AI which beat him last time, and he'll simply do something he would never normally do to throw off the computer. It's a difficult enough problem that in pretty much all commercial computer games with a fog of war feature, the computer is just programmed to cheat by ignoring the fog and seeing everything.

"teaching itself"

[Bobtree]

> DeepMind learned to beat the best go players in the world by teaching itself through trial and error.

AlphaGo was trained on databases of historical games. It looks for moves that are similar to what a human pro would play, and then reads out sequences to score the strength of the resulting position. It did not learn by itself from scratch. Once proficient, it was played against itself to improve.