Clan Crazy Forum

    • Register
    • Login
    • Search
    • Categories
    • Users
    • Groups
    • Search

    OpenAI v. OG

    DOTA2
    2
    3
    72
    Loading More Posts
    • Oldest to Newest
    • Newest to Oldest
    • Most Votes
    Reply
    • Reply as topic
    Log in to reply
    This topic has been deleted. Only users with topic management privileges can see it.
    • vaette
      vaette Crazy last edited by vaette

      Somewhat likely I am actually the last to watch these, but OpenAI (the deep learning system trained on DOTA) played OG a bit back, and where their previous games against Wings (I think?) at TI looked pretty weird I think they have tweaked it enough that the games have a look of normal play now:

      Game 1:


      Game 2:

      Still a very limited meta and no doubt a lot of weirdness, but it is if nothing else interesting to see what skill and item builds the AI arrives at (because it does learn all that stuff independently). E.g. no early point in Nethertoxin on Viper.

      1 Reply Last reply Reply Quote
      • B
        Black Bob Crazy last edited by

        Watching the AI say we have a 95% chance of winning is just epic, when they can play with 100% of the hero pool I will be impressed!

        1 Reply Last reply Reply Quote
        • vaette
          vaette Crazy last edited by

          Unfortunately they are dropping doing DOTA bots with this showcase. To some extent I, as someone who does work in the shadows of these things, expect it is because they haven't made headway with the core issues the bots have. To some extent what they do well was already on display with the SF mid bot and has only been scaled up without being changed in substance.

          I.e. they don't understand anything but early-game push, notably they had to be coaxed ridiculously to value Roshan (they had a whole subsystem to give Roshan low health randomly in training to make the AI see any value in committing to the kill), as they could not figure out the value of getting an aegis when the use of the aegis is a passive event minutes (and thus hundreds of thousands of moments by their reckoning) in the future. Also it is pretty notable that they don't really get supporting (again a hardcoded system just forces whoever is poor to buy wards), and heroes like e.g. Oracle, Dazzle, and Shadow Demon, with dramatic saves, are not included in the list of heroes. With classical AI these would have been excellent (as after all situation judgement and complete selflessness are a given with classical AI), but I expect that it was very hard to with reinforcement learning, i.e. neural nets, actually get the AI to understand momentary sacrifice for a greater good.

          1 Reply Last reply Reply Quote
          • First post
            Last post