Cake
  • Log In
  • Sign Up
    • My professor at Stanford said there is simple mathematics behind judging a Phd dissertation: its value is inversely proportional to the number of pages it contains. The pages increase as the quality of writing decreases; the author fills it with air if he or she was covering for not having advanced science much.

    • From my, admittedly limited, understanding of the process, the way AlphaZero did it was a two phase process. First phase consisted of playing a bunch of games against itself, making completely random (but chess legal) moves. Second phase consisted of letting a machine learning algorithm work on finding patterns in moves leading to victory, similar to the way neural networks train to find images of cats. It ended up with atrained model that doesn't have to brute-force examine all the possible moves (like Stockfish does) - it can dismiss vast number of possible moves easily (none of them ever led to winning games).

      But the crucial part in its 'out of this world' style of play was the random nature of moves in the first phase. Its the reason why it avoids all the common styles of play humans have - humans all learned from other humans. AlphaZero didn't, so it gave every move a fair chance and ended up with lots of winning moves humans would outright dismiss because "that's not the way you play chess".

    • this stuff is fascinating.

      I wonder how Alpha Zero would tell soccer players to build an attack or inform Bill Belichick what defense to run against each oppositions formation and the strength and weakness of their opponents.

      Would Alpha do something crazy like put 8 players on one side of the soccer field or have no striker (is that like giving up a pawn?)

    • That's where we bump against the limitations of this approach. Soccer (or any other real-world games) has immeasurably more play states than chess, go, or other board games with discrete positions (how many positions are there for a soccer player?). Not only that, but the players' ability is wildly different (and changes over time!). Modelling such games in a fashion that would be usable for AlphaZero would be extremely hard. Besides, you would even have trouble with the first phase of the process: it would be really hard to find two teams willing to play a couple of billion of games with each player doing totally random moves. :-)

    • sounds like a perfectly reasonable human response (I was thinking the exact same thing)...

      maybe Alpha Zero would say no problem let's just rule out 90% of the actions that we know won't result in a win like putting 10 players on one side of the field etc. It probably would start with the standard framework coaches start with and "learn".

      Could Alpha Zero "watch" games by analyzing the movements of players from previously recorded games over the past 10 years? I'm sure they can turn the players movements into some kind of data that might look a bit like chess except all the players are the queen that can move in any direction?

    • Actually, that's an interesting idea. Place GPS devices on the players and the ball with some player metrics like weight and height into a data stream and feed that data to AlphaSoccer. In the past that would have been called statistics and people like Billy Beane would have become famous for it in MoneyBall.

      Baseball has always leant itself to statistics, but if you had player and ball position data in soccer plus AI, whoa.

    • Yes, it can analyse the recording of existing games, but that would only get us up to AlphaGo, which learned Go by analysing human played games. The real breakthroughs come from forgoing the human experience and practice in particular game and starting from tabula rasa.