Go, Starcraft and AI

MONDAY, MARCH 21, 2016

It's been approximately a week since the showmatch between DeepMind's AlphaGo and Lee Sedol came to a close. As someone who recently started playing go (around a year ago), I couldn't help but take some interest in the game. AlphaGo won 4-1, marking a moment that people thought would be coming a decade from now.

This was my first time watching a televised go match, and I have to say that the official DeepMind stream, as well as the AGA stream both did a great job of explaining what was happening. Whenever I thought a play was insane, one of the 9d pro commentators would explain how reasonable of a response it was.

One of the things that was mentioned throughout the go matches was that DeepMind's system can be applied to a variety of games, and in fact it had been playing over 30 different old school arcade games as well. The next game that they might focus on though, is Starcraft. While I do think that we can get to a point where a go AI will consistently beat humans, I'm not so sure about Starcraft.

Of course, there are some caveats to that. First, the AI's APM and eAPM will have to restricted to a human level. Second, while I do think we can get to a point where AI can beat Life, Maru, INnoVation etc. in series, I don't think they'll be able to consistently beat them 4-0. Starcraft is not a game of perfect information, and it has a metagame. So any strategy that the computer picks could be punished by an all-in, or edged out by a greedier build.

Much of this depends on the level of micro of the human and the AI, which is easy to conflate with battlefield strategy, something harder to evaluate. Even with APM restrictions, the AI could presumably be better than a human player at microing their army, which would slow erase any potential metagame advantage. Maybe the end goal is to beat pro players in as little APM as possible, where APM cap is somewhat analogous to giving stones in go. In the AlphaGo match, I would have been curious to see a game in which Lee Sedol had a stone handicap.

If DeepMind really does decide to pursue Starcraft, I'm excited to see what they can come up with, and at what APM thresholds before they can win. Google, make it happen, for science.