Considering how much ai is hyped it is getting harder and harder not to be a skeptic.
Is AlphaZero really a breakthrough in AI?
21–30 of 84 posts
Re: Is AlphaZero really a breakthrough in AI?
#22Whether it was a breakthrough or not, I have to say, the moves it played were certainly "creative" in a profound sense.
Re: Is AlphaZero really a breakthrough in AI?
#23The article glosses over why the 4 hours was possible. Firstly, a major challenge in training an AI of this sort is getting enough labelled data. They played 300,000 games from memory. Under normal circumstances, that requires access to 300,000 games played by experts so the AI can learn to copy what the export does. That is how Alpha Go did it. AlphaZero neatly side steps this by generating it's own training data by…
Forgive me if this I'm being waaay too naive, but it doesn't seem like the parallelism side of this is a big deal. Playing N games in parallel and then feeding them back in for training doesn't seem like an incredibly innovative idea? On the timeline issue of course, I don't it's reasonable to make any assumptions about what that implies about the algorithms. It may be that Chess was solved, the dev team went on a 6…
Playing the games simultaneously is actually innovative in how it let's them learn faster. If you play one game at a time consecutive moves are highly correlated which makes it difficult to make learning updates after each move. If you're playing 1250 parallel games that are starting and restarting so that they're at all at different stages of play, you can do an update after every move in every game and make better progress faster. This is known as Asynchronous Advantage Actor-Critic (A3C) and is a highly cited method discovered by the DeepMind team [1].
Re: Is AlphaZero really a breakthrough in AI?
#24Earlier quoted context omitted.
It could perhaps be said, then, that the Alpha series as a whole is a rolling breakthrough? Each new generation introduces changes that do not seem major compared to the ideas in the foundation of Deep Learning, but this actually allows to surpass state-of-the-art, improving performance on all relevant parameters (and not by a few percent). Dismissing these updates as technicalities is a dubious position.
Isn't that how all newer version of chess engines - including stockfish, improve? Edit: https://github.com/glinscott/fishtest http://tests.stockfishchess.org/tests
Prior to AZ, some competent people had doubts if self-taught NN could compete at all with a mature chess engine at its strongest setting. You know how that went. The improvement rate of subsequent Alpha revisions is incomparable to what software like chess engines shows. Hence, the tendency to evaluate each one as a breakthrough.
Personally I believe that "breakthrough" is a non-technical word that's hard to define, but "improvement rate surpassing most optimistic expectations for a given field at a given time" could probably qualify as a criteria.
Re: Is AlphaZero really a breakthrough in AI?
#25Whether it was a breakthrough or not, I have to say, the moves it played were certainly "creative" in a profound sense.
If possible at all, could you give an example for casual chess players?
Re: Is AlphaZero really a breakthrough in AI?
#26Earlier quoted context omitted.
It would be good to see a definitive playoff. I've no real doubt that AlphaZero would triumph, but in people's breathless coverage of the games nobody seems to point out Stockfish's various 5-10 pawn blunders (all of which my version of Stockfish finds when annotating the games with 60 seconds per move).
What do you mean by "various 5-10 pawn blunders"? Is that a count of the number of blunders or some sort of score?
Stockfish and other mainstream engines will judge a position in units of centipawns - one hundredth of a pawn. Largely this will be to do with material on either side. For example, if you throw away a pawn with all else being equal, I'd be up 100 centipawns. If I'm white this gets written as "+1.00", and if I'm black it's "-1.00". The score is also based on a strategic evaluation of the position (and positions to come) using (in Stockfish's case) some heuristics. It's worth noting that it's this evaluation of a position at which AlphaZero appears to be three orders of magnitude better than Stockfish.
In some of the games vs AlphaZero, Stockfish makes errors that it _itself_ appears to be able to judge as huge blunders. In game 3 (which people view as a masterpiece of long term strategic thinking by AlphaZero) one of these is at least a +10 swing to AlphaZero as white. That's about the same as throwing away your queen. Without the weird time controls put in place, it seems unlikely that we'd be seeing blunders like that. As I said before, I'd still expect AlphaZero to win, and in many cases it was already ahead before these mistakes, but it's worth mentioning in any analysis.
Re: Is AlphaZero really a breakthrough in AI?
#27The article glosses over why the 4 hours was possible. Firstly, a major challenge in training an AI of this sort is getting enough labelled data. They played 300,000 games from memory. Under normal circumstances, that requires access to 300,000 games played by experts so the AI can learn to copy what the export does. That is how Alpha Go did it. AlphaZero neatly side steps this by generating it's own training data by…
> AlphaZero neatly side steps this by generating it's own training data by playing itself. If how to do this was "obvious", it would have been done a long time ago. Learning by self-play is nearly as old as AI itself. TD Gammon, one of the very first algorithms to reach superhuman levels in a nontrivial game, learned by self play. The basic ingredients for AlphaGo Zero, Monte-Carlo tree search and use of a convnet to…
its own training data
Re: Is AlphaZero really a breakthrough in AI?
#28Whether it was a breakthrough or not, I have to say, the moves it played were certainly "creative" in a profound sense.
If possible at all, could you give an example for casual chess players?
Seconded. We all want a glimpse of what the "AI" is doing, and see that one step ahead that human minds could never do :)
Re: Is AlphaZero really a breakthrough in AI?
#29The article glosses over why the 4 hours was possible. Firstly, a major challenge in training an AI of this sort is getting enough labelled data. They played 300,000 games from memory. Under normal circumstances, that requires access to 300,000 games played by experts so the AI can learn to copy what the export does. That is how Alpha Go did it. AlphaZero neatly side steps this by generating it's own training data by…
> AlphaZero neatly side steps this by generating it's own training data by playing itself. If how to do this was "obvious", it would have been done a long time ago. It was done a long time ago. Arthur Samuel's chequers program (1959) improved by playing against itself in thousands of games. People can and do improve that way too: https://www.chess.com/forum/view/general/playing-against-you...
> People can and do improve that way too
When learning Chess or Go, playing whole games with myself would be too tedious. Do people really do this?
But I have played countless opening sequences in chess and many many joseki lines in go, I have practised various piece mating combos in chess and different yose techniques in go, one does chess problems from newspapers or books and of course everyone knows you do go problems (tsumego) to get strong at go – this means I only ever play small subsets of both games against myself. Never whole games. It is this ability that the machines have over us.
The only whole games I've played are both sides of grandmaster level competitive-games in each board-game. But then, that's not really playing against myself where I choose each move. Some expert has chosen the move for me. Interestingly, AlphaZero does not need to learn this way because it can play itself so many times. Because that amount of times would be too tedious (possibly even unfeasible) for a human we will always continue to learn from expert play.
When AI is smart enough to know what expert play _is_ it may go back to using that technique.
Re: Is AlphaZero really a breakthrough in AI?
#30Earlier quoted context omitted.
> AlphaZero neatly side steps this by generating it's own training data by playing itself. If how to do this was "obvious", it would have been done a long time ago. It was done a long time ago. Arthur Samuel's chequers program (1959) improved by playing against itself in thousands of games. People can and do improve that way too: https://www.chess.com/forum/view/general/playing-against-you...
> People can and do improve that way too When learning Chess or Go, playing whole games with myself would be too tedious. Do people really do this? But I have played countless opening sequences in chess and many many joseki lines in go, I have practised various piece mating combos in chess and different yose techniques in go, one does chess problems from newspapers or books and of course everyone knows you do go prob…
> When learning Chess or Go, playing whole games with myself would be too tedious. Do people really do this?
Chess World Champions Robert Fischer and Magnus Carlsen notoriously did as children. Magnus Carlsen even recommended it as one of his top 13 tips to improve :