If you have a new account but are having problems posting or verifying your account, please email us on for help. Thanks :)
Hello all! Please ensure that you are posting a new thread or question in the appropriate forum. The Feedback forum is overwhelmed with questions that are having to be moved elsewhere. If you need help to verify your account contact

DeepMind AI battles vs. top level SC2 players

  • 30-01-2019 7:08pm
    Moderators, Category Moderators, Arts Moderators, Business & Finance Moderators, Entertainment Moderators, Society & Culture Moderators Posts: 18,253 CMod ✭✭✭✭

    For those not aware DeepMind is an AI developed to play Starcraft 2 and improve AI research. Unlike most other AI games before SC2 is a game with incomplete information real time strategy which means the AI needs to adapt constantly to what it learns (unlike for example chess where it can see all pieces at all times). In December it went up against top level pro players TLO and MaNa (they commonly play at global tournaments and have earned over $300.000 in price money between them with MaNa seen as a top 10 player for his race globally).

    There are spoilers how it went in the article above but if you want to see the games with commentary you can find it here:

    All I'll say is that this has some great promise for years to come...


  • Registered Users Posts: 20,558 ✭✭✭✭dreamers75

    But how did the AI question his parentage, wonder if his mother was promiscous or even question his ethnicity?

  • Moderators, Computer Games Moderators, Sports Moderators Posts: 14,526 Mod ✭✭✭✭Darkglasses

    The TLO match was pretty pointless really. The build of Alphastar was fairly poor, and TLO (not a top player for the last few years) was playing his weakest race. Any mid GM player playing their best race would have beaten that build of Alphastar.

    The build that played against Mana looked legit, and very convincingly beat him in the best over 5 games. However, the micro-control Alphastar was doing was undoubtedly superhuman, which slightly defeats the point of the project. Nobody has any doubt that an unrestricted high-quality AI could beat a human, easily. It's interesting when it's given the limitations that humans have.

    I also thought it was a bit pointless showing an AI that could only play Protoss vs. Protoss on one map several times over. That's not how competitive SC2 works at all. Considering how much Alphastar improved in the space of a week between the TLO and Mana games, they really should have waited a small bit longer to overcome the map variety and race variety that is part and parcel of the game. It would still definitely win, given enough computing time.

  • Moderators, Social & Fun Moderators Posts: 28,633 Mod ✭✭✭✭Shiminay

    The complexity of learning another map and race though is exponentially higher for the AI though - it's a LOT more complex, but I'm sure it's part of the plan. The TLO matches were to help it "learn" and it costs TLO nothing in terms of reputation to do that because we all know he's no longer a top tier player. This is a marketing partnership for Liquid too, so they're clearly playing ball with Google on it :) It's not necessarily about showing that an AI can compete in a competition, but that it can learn a very complex game like SC2.