Hex AI Discussion and Theories

    • Hex AI Discussion and Theories

      With the recent Siege announcement I've seen a lot of talk about the programs in the background of Hex that plays our PvE opponents in the game.

      There was a vice article that describes some of the ideas.

      google.com/amp/s/motherboard.v…article/8q8x7g/hex-tcg-ai

      In the article it was mentioned that:

      "The AI in Hex isn't just playing the game, it's also watching and learning. Every one of the millions of games already played have been recorded by the AI, which is aggregating all of that information to better understand how the game is played. When certain card combinations become popular, the AI will begin to anticipate and prepare to counter them."

      Now, I wondering if some of the 'dumb' AI moves we see (and potential Siege Defenders are worried about) are because a vast majority of Hex players will misclick, misplay, or might not be that good at the game.

      I'm sure there must be some fail-safes in place to prevent the AI from learning how to play badly but that seems like it would be a tricky bit of design.

      The example that took me down this line of thought was that, with the most recent Set, the AI has begun self-milling in encounters where it uses Mill decks. Which is a very stupid play, barring specific circumstances.

      However, I also know there was a recent 'Self-mill' deck for FRA runs. What if the clever Hex AI saw enough people running the self-milling decks enough times that some switch was flicked and the AI now 'knows' that self-milling leads to victory?

      Basically, are we as a community so bad at this game we made Chris Wood's AI dumb? :P
    • I love the idea that the AI self-mills now due to the proliferation of Succumb decks and reanimator decks... Silly AI. It tries so hard. :D

      Really, even if that's not the case, I actually think the Hex AI is pretty solid for what it is. I'm not too concerned with it running a majority of what I have in mind well enough. Dino ramp will still make those pesky attackers bitterly regret the day they showed up at my castle. If it can't figure out some my wackier HC or Scheme decks, that's okay. I recognize it's a bit much to ask.
    • Darklight wrote:

      The example that took me down this line of thought was that, with the most recent Set, the AI has begun self-milling in encounters where it uses Mill decks. Which is a very stupid play, barring specific circumstances.


      However, I also know there was a recent 'Self-mill' deck for FRA runs. What if the clever Hex AI saw enough people running the self-milling decks enough times that some switch was flicked and the AI now 'knows' that self-milling leads to victory?

      Basically, are we as a community so bad at this game we made Chris Wood's AI dumb? :P
      I don't think training an AI based on HOW your opponent plays his cards make sense. It should be trained on how the AI plays his own cards and see if that leads to a victory or loss. Of course, the cards of the opponent plays need to be involved in the calculation of which card to play next as AI, but not the way HOW he plays them. The reason a deck self-mills would be for an entirely different reason than why the AI mills his opponent with the same card. Instead, an AI that is self-milling and therefore loses all the time should eventually be able to recognise that self-milling doesn't work, and shouldn't even consider that his opponent DOES self-mill to victory.
      That being said, with skewed matches where the human wins 95% of the time, it will be hard for the AI to train a certain play-style or pattern that often leads to victory, since the AI so many times, even when he's playing with different strats each game. So it might be true that because self-milling is a thing now, the AI will try it out, but can't figure out WHY it leads to a loss, because he loses most of the other games too.

      That's my perspective on this, with some limited knowledge of machine and deep-learning.