The Tao of Gaming

Boardgames and lesser pursuits

More DeepMind — This time Starcraft

I’ve never played SC (etc), so watching these replays isn’t very informative. But still interesting. If you want to skip to the first replay, its at about minute 44 (after 30 minutes of count down and 14 of discussion). The second match starts at around 1:32. If you watch one match, G4 of the second match at around 2h mark.

Here’s the main article

A few thoughts (mentioned in the video):

  • The AI (‘AlphaStar’) actually “clicked” less than the PROs (200-300 “Actions per minute” vs 500-600 for the human). That being said the commentators kept praising AlphaStar’s micro-management (“micro”) in terms of ability to maneuver in battles (for example —  retreat a single almost dead unit from a battle while having the rest press) was praised and may have actually been one of those precision benefits that humans just can’t match.
  • In the 5 game match, they actually fielded five different agents (with order randomized). They used a semi-guided genetic strategy then tried to minimize ‘exploit-ability’. The human was not aware of this (until after the fact), which may have affected strategy. After a game the human “adjusted” but was then facing a totally new strategy.
  • This wasn’t Lee Sedol where they were playing the undisputed, but Starcraft is a much more complex (and real time) game, so impressive.




Written by taogaming

January 25, 2019 at 5:17 pm

3 Responses

Subscribe to comments with RSS.

  1. Starcraft is so mechanical that I don’t find this impressive.

    Iain Cheyne (@Icheyne)

    January 25, 2019 at 6:55 pm

    • Controlling it sure, but understanding the board position in real time is no small feat. It’s not nearly as hard as a computer vision problem (the API identifies all the visible objects, I assume), but if robotics and CS have taught me anything, its that things that we think are hard are often trivial for computers (giant arithmetic problems) and things that a toddler could handle are often impossible for computers.


      January 26, 2019 at 10:41 am

  2. Thanks! I haven’t played StarCraft at all, though some of the people I know have played it, but it’s another fascinating look at how AI is being used to learn to play games better.

    Chess and Go strategy look a lot more like pure math (even if DeepMind uses neural nets) than Starcraft.

    Eric Brosius

    January 25, 2019 at 11:05 pm

Comments are closed.

%d bloggers like this: