AlphaZero plays games with (1) perfect information and (2) well-defined winning conditions . Neither of these hold for most human-learning scenarios. I can imagine that a healthy dose of probability theory (and probably more advanced stuff I don't know about[1]) might improve (1), but (2) is going to keep computer scientists and philosophers and ethicists arguing for quite a long time. :) [1] get the joke, eh? eh? eh…
> AlphaZero plays games with (1) perfect information
I'm not sure why this matters? Everyone plays chess with perfect information. Both players see the entire board and all possibilities unlike, say, Scrabble or poker.
I would submit that we already have an example of self-play being used as part of a strategy to learn chess: chess problems.
Something like "Here's a board position. It looks utterly hopeless but the problem says "Black to mate in 7 moves". How can you get there from here without relying on White making any beginner's mistakes?" is pretty much self-play.
AlphaZero plays games with (1) perfect information and (2) well-defined winning conditions . Neither of these hold for most human-learning scenarios. I can imagine that a healthy dose of probability theory (and probably more advanced stuff I don't know about[1]) might improve (1), but (2) is going to keep computer scientists and philosophers and ethicists arguing for quite a long time. :) [1] get the joke, eh? eh? eh…
> AlphaZero plays games with (1) perfect information I'm not sure why this matters? Everyone plays chess with perfect information. Both players see the entire board and all possibilities unlike, say, Scrabble or poker.
I think GP meant that in the sense of "AlphaZero can only play games that have perfect information". It's a restriction of the algorithm, not a statement about how AlphaZero approaches the games it plays.
This is why AlphaGo leveled up into AlphaZero playing Chess, and didn't learn to play Starcraft (yet).
Short answer: No, there's nothing new here that can inform better human learning. Longer answer: The concept of self-play isn't new in any sense. All chess players use this technique to some degree. None use only this technique. The advantage of self play is that there's no risk of accidentally picking up someone else's an incorrect assumption. Since you're deriving everything from scratch. Some people take this to e…
>unfortunately I'm blanking on his name
John Nash (supposedly) had this mindset? Is that who you're thinking about?
> AlphaZero plays games with (1) perfect information I'm not sure why this matters? Everyone plays chess with perfect information. Both players see the entire board and all possibilities unlike, say, Scrabble or poker.
I think GP meant that in the sense of "AlphaZero can only play games that have perfect information". It's a restriction of the algorithm, not a statement about how AlphaZero approaches the games it plays. This is why AlphaGo leveled up into AlphaZero playing Chess, and didn't learn to play Starcraft (yet).
Short answer: No, there's nothing new here that can inform better human learning. Longer answer: The concept of self-play isn't new in any sense. All chess players use this technique to some degree. None use only this technique. The advantage of self play is that there's no risk of accidentally picking up someone else's an incorrect assumption. Since you're deriving everything from scratch. Some people take this to e…
>unfortunately I'm blanking on his name John Nash (supposedly) had this mindset? Is that who you're thinking about?
That wasn't who I had in mind, but thanks for sharing that example. I think the guy I'm thinking of is at Cornell and still alive. He also might actually be in CS instead of Math. I tried googling it but, unfortunately, "math professor who doesn't read papers" didn't come up with any results.
My impression of Plato is he channeled different people/characters in his writing in order to create adversarial conditions in which he could improve his rhetoric. Perhaps this is similar to AlphaZero's technique?
>unfortunately I'm blanking on his name John Nash (supposedly) had this mindset? Is that who you're thinking about?
That wasn't who I had in mind, but thanks for sharing that example. I think the guy I'm thinking of is at Cornell and still alive. He also might actually be in CS instead of Math. I tried googling it but, unfortunately, "math professor who doesn't read papers" didn't come up with any results.
Shinichi Mochizuki developed a new theory (IUT) which eventually yielded a proof to the abc conjecture. I believe he largely developed it in isolation and thus when it was published it took years to bring the rest of the community up to speed. Not sure if he doesn’t read other’s papers but maybe this is what’s you were thinking about?
AlphaZero plays games with (1) perfect information and (2) well-defined winning conditions . Neither of these hold for most human-learning scenarios. I can imagine that a healthy dose of probability theory (and probably more advanced stuff I don't know about[1]) might improve (1), but (2) is going to keep computer scientists and philosophers and ethicists arguing for quite a long time. :) [1] get the joke, eh? eh? eh…
> AlphaZero plays games with (1) perfect information I'm not sure why this matters? Everyone plays chess with perfect information. Both players see the entire board and all possibilities unlike, say, Scrabble or poker.
I disagree that it plays with perfect information, in that it is playing against a different version of itself. It can "discover" that it is playing with perfect information, i.e. learn strategies that are more effective against a clone of itself, but it first has to develop some "sense" that this is the actual case.. and throwing in some random stock fish opponents or opponents that are clones of itself, but cloned at different stages of thier learning process, would remove this "perfect information" caveat. Unless perfect information only applies to the state of the board.. in which case, perhaps a "fog of war" visibility algorithm could push that boundary.. but then your talking about self learning with a blindfold type scenario.