As a scientist I know that, in any comparison, you eliminate as many variables as possible. Limiting SF to less thinking time than AZ stands out alone as a boldface attempt to give AZ the advantage.
...
Pretty sure both sides had 1 minute per move.
My view is that by human standards it could generate a good opening book (AlphaZero would build a significantly better one!). The body of very strong computer games might be used to generate a better book. Of course to do so, they need to be given free rein some time in the opening!
Agreed, but that's not the point of contention. The issue appears to be that Stockfish, as it was used, was not used in a way that maximized its winning potential. It was not the best version of Stockfish, it was a recent version of Stockfish, unoptimized. AlphaZero, by comparison, was optimized and running to special hardware designed specifically for the task.
I don't mean to diminish AlphaZero by this. It's an amazing program! And the paper is very up front about what parameters were used to test with. You can't, however, that it beat Stockfish because what it beat was a version of Stockfish that was not optimized or "tweaked" to its full advantage. That is to say, you wouldn't have run AlphaZero in a similar fashion... on hardware that it's not optimized for.
Long story short, I'd like to see a rematch with Stockfish configured by someone who wrote it.
AlphaZero: "But, I've got a job to do, too. Where I'm going, you can't follow. What I've got to do, you can't be any part of, Chess. I'm no good at being noble, but it doesn't take much to see that the problems of two little chess engines don't amount to a hill of beans in this crazy world. Someday you'll understand that."
AlphaZero: "But, I've got a job to do, too. Where I'm going, you can't follow. What I've got to do, you can't be any part of, Chess. I'm no good at being noble, but it doesn't take much to see that the problems of two little chess engines don't amount to a hill of beans in this crazy world. Someday you'll understand that."
I'll have to comment in some code about beautiful friendships in post production.
So I'm waiting............................... As I predicted. No more games. I wish I was wrong. I'd like to see some more games.
So I'm waiting............................... As I predicted. No more games. I wish I was wrong. I'd like to see some more games.
It looks like the Google people are taking their cue from IBM.
So I'm waiting............................... As I predicted. No more games. I wish I was wrong. I'd like to see some more games.
It looks like the Google people are taking their cue from IBM.
Yes that is what I was trying to say with this thread. The DeepMind team used chess and Stockfish to show what could be done. They are not doing this for chess. Chess was a ruler by which they measure. Done. Moving on. Do not expect to see them again in this space. I hope they can apply it to medicine, and science.
So I'm waiting............................... As I predicted. No more games. I wish I was wrong. I'd like to see some more games.
It looks like the Google people are taking their cue from IBM.
Yes that is what I was trying to say with this thread. The DeepMind team used chess and Stockfish to show what could be done. They are not doing this for chess. Chess was a ruler by which they measure. Done. Moving on. Do not expect to see them again in this space. I hope they can apply it to medicine, and science.
...except that they misrepresented what they had actually done. You can bet that if and when AlphaZero can defeat Stockfish soundly under real match conditions, they will do it.
Having watched the Netflix AlphaGo documentary where one of the key team members talks about how he was trained in chess from a young age, it is quite easy to believe that the premature release of private test results as if there were a public match with a complicit Stockfish team has not only a motive of free publicity at Stockfish's expense, but also possibly an ego boost for said team member. Maybe he told his higher-ups he'd beat Stockfish in 2017 but could not without tinkering with the settings, maybe his annual bonus depended on it...the point is, who knows? It all happened in private where it should not have (not if it was going to be released to the press).
it was NOT a match. It was a test if A0-Go method can be generallized so that it can learn to play a game with no game specific knowledge. At that has been proven. Whether result is bit stonger than SF or bot weaker does is no consequence for answering actual research question
Well put @petrip.
it was NOT a match. It was a test if A0-Go method can be generallized so that it can learn to play a game with no game specific knowledge. At that has been proven. Whether result is bit stonger than SF or bot weaker does is no consequence for answering actual research question
Not the case. The AI is already "generalized", in that it was not developed specifically to play Go or any other single game or even genre of games. If you mean that they wanted to see if AlphaZero could teach itself GM/engine level chess without a shred of "assistance" as given to AlphaGo along the way by the team, then fine. Please try to be more accurate.
As for the "test" assertion...preliminary and/or private test results are generally not plastered all over the news.
As a scientist thinking in terms of the scientific method, there's not much anyone can claim to know about AZ's strength compared to SF.
Many people here are stretching the truth or making wild guesses in trying to defend one side of the other. Some claims, like the current version of SF is 8 is clearly wrong - I've been analyzing my games with SF 8+ for a couple months.
As a scientist I know that, in any comparison, you eliminate as many variables as possible. Limiting SF to less thinking time than AZ stands out alone as a boldface attempt to give AZ the advantage.
They didn't "limit Stockfish". They provided Stockfish with an extremely powerful machine on which it could run. Stockfish simply can't run on the hardware (TPUs) on which AlphaZero ran (this hardware is now available online as a new option on the google cloud computing service, incidentally, aimed principally at deep learning developers): these are designed for the sorts of calculations neural nets need to do (which are principally matrix operations on large numbers of elements in parallel). They provided AlphaZero with the hardware that was essential for it to achieve its potential: the neural network it used was clearly quite large.
Note that anyone could try to beat Stockfish on a decent machine as convincingly using any engine (including another version of Stockfish!) with as much computing power as they desired. No-one has even seriously claimed they could do as well as AlphaZero did.
Consequently, I hope AZ leads us into a new style of computer analysis that may lead to something like computer analyses of our games telling us, "You should not have tried a K-side attack: you should have posted your N on the excellent c5 outpost and then pushed your d- and e- Pawns up the middle."
But the game was done under such poor conditions, maximizing variables that can't be compared, so for me there's nothing to assume or argue about one way or the other.
Actually, AlphaZero found chess boring and is now into writing poetry, for Deepmind team desperation.
There is some truth in that, but the problem is that opening books are less and less effective for even very powerful engines, never mind AlphaZero. All they can really do is avoid some disasters and defer the real battle. The very best an opening book can ever do is to use previous games as samples to use in analysis (which relies on trusting that the games are adequately accurate). For example, this can go badly wrong if you are playing the French against AlphaZero, even if it might be ok against a GM!
I disagree. Opening books can be invaluable because they are based on years of experience that the computer doesn't have to do itself using up plys, CPU cycles and game time.
Think of it as nearly infinite plys by humans codified into opening moves. Again, in the relatively few games that AlphaZero won (the vast majority were draws), it was because Stockfish effectively blocked itself in with its opening.
Have you seen the games? SF didn´t have any problem in the opening, but went down in the middlegame, how many times must it be said? In addition, those "years of experience" are hilarious for an engine.
There is some truth in that, but the problem is that opening books are less and less effective for even very powerful engines, never mind AlphaZero. All they can really do is avoid some disasters and defer the real battle. The very best an opening book can ever do is to use previous games as samples to use in analysis (which relies on trusting that the games are adequately accurate). For example, this can go badly wrong if you are playing the French against AlphaZero, even if it might be ok against a GM!
I disagree. Opening books can be invaluable because they are based on years of experience that the computer doesn't have to do itself using up plys, CPU cycles and game time.
Think of it as nearly infinite plys by humans codified into opening moves. Again, in the relatively few games that AlphaZero won (the vast majority were draws), it was because Stockfish effectively blocked itself in with its opening.
Have you seen the games? SF didn´t have any problem in the opening, but went down in the middlegame, how many times must it be said? In addition, those "years of experience" are hilarious for an engine.
Yes, I've seen the winning games. They are positional nightmares for Stockfish... but all are predicated on bad openings. The middle games, were AlphaZero man handled Stockfish in those wins, are available as wins because Stockfish, IMHO, didn't have its opening book.
Having said that, it's still a remarkable feat to feat Stockfish with a Neural Network that was trained in a trivial amound of time.
As an enthusiast of Chess I would very much like to see a match where both sides felt their programs were tuned and played on the best available hardware. I don't think that's too much to ask...
[And let's not forget... the vast majority of the games played were draws. By vast, I mean more than you would play in an entire GM tournament.]
"Having said that, it's still a remarkable feat to feat Stockfish with a Neural Network that was trained in a trivial amound of time."
Given the hardware involved, the "time" involved as far from trivial. The processing power involved in those 4-some-odd hours would have represented years/decades of a desktop computer's "time" depending on where you place the bar/threshold for a desktop computer.
The 4 hours would not be a triumph of AlphaZero in terms of machine learning or AI, that's a triumph of hardware over inferior hardware 
.
Here is a recently released, but older video presentation regarding AlphaZero that I found interesting:
https://www.youtube.com/watch?v=Wujy7OzvdJk
Nope, method of learnign games was not generalized. Yed DCNN have been used for mab´ny things but here they took all the go specific information out and used only a module that can tell legal moves and outcome. MCTS has not really used outside Go at all. So combination of DCNN and MCTS on domain of several board games is new.
As for preliminary results: no these are final results. Result was the method of A0-Go can be used also in other board games. and enough of information is given that other can re-produce and verify the results. In go community there is a project that reached a reasonable strength and just about anyone with proper GPU can contribute
Furthermore there would be other games that are hard for computer like Amazons but thats again for someone else to do.
"here they took all the go specific information out and used only a module that can tell legal moves and outcome."
...which...say it with me now...would be considered "generalized". The machine learns from scratch by being given the bare minimum in terms of boundaries/ruleset and successful/unsuccessful outcomes without human assistance so as to prove the learning is "pure" and no human biases/valuations are present. At least, that should be the ideal goal.
You can't have your cake and eat it, too.
There was a high degree of generalisation in the approach used by DeepMind, to the extent that they used the same choices of key parameters for the learning process: minibatch size, learning rate schedule, number of batches used.
Chess was actually the least computationally demanding of the three games with these parameters (because it is the third most complex of the games in terms both of numbers of legal moves and typical length of games). An interesting probably associated result was the strong monotone relationship between the complexity of the game and the Elo achieved by the AlphaZero learning. In chess it was "only" about 100 points stronger than Stockfish, in Shogi it was several hundred points stronger than the top rated opposition and in Go it was a stunning more than 1000 points stronger, despite the very deep development of this game by professionals over its very long history. (Of course among computers, by far the strongest opposition was the previous AI developed by DeepMind, AlphaGo!)
The original contention of this thread was that Google Deep Mind would quickly move onto new topics, rather than hyper-analyze and redo the StockFish match.
To that end, in today's FTimes, US edition, the front page has a story on Google-DeepMind, London hospitals, and spotting eye diseases using AI on eye scans.
QED? Maybe.
As a scientist thinking in terms of the scientific method, there's not much anyone can claim to know about AZ's strength compared to SF.
Many people here are stretching the truth or making wild guesses in trying to defend one side of the other. Some claims, like the current version of SF is 8 is clearly wrong - I've been analyzing my games with SF 8+ for a couple months.
As a scientist I know that, in any comparison, you eliminate as many variables as possible. Limiting SF to less thinking time than AZ stands out alone as a boldface attempt to give AZ the advantage.
Consequently, I hope AZ leads us into a new style of computer analysis that may lead to something like computer analyses of our games telling us, "You should not have tried a K-side attack: you should have posted your N on the excellent c5 outpost and then pushed your d- and e- Pawns up the middle."
But the game was done under such poor conditions, maximizing variables that can't be compared, so for me there's nothing to assume or argue about one way or the other.