Everything AlphaZero

Yesterday, I was just casually checking my email and news feed. And it was at that moment that I saw words that spelled what seemed to be unfathomably impossible i.e. "Google's AlphaZero Destroys Stockfish In 100-Game Match." I was so puzzled, curious, upset, and excited all at the same time. I went to the article, and was surprised. At 100 games, Stockfish could not even win a single game. AlphaZero won the match +28-0=72. 28 wins and zero losses against the same exact chess engine that won the TCEC and Chess.com event/tournament. What!? How? Even though the article answered questions most of the readers were wondering, e.g. what openings were used, what were the time controls, what method was used, etc... But I still have a plethora of questions [I literally have about at least 200 total questions but I don't want to list them all numerically and exhaustingly]. So I will ask just 12 basic, questions that I openly invite any person to participate in. Also, ask any questions that have not been asked or state any answers that have not been stated. Please feel free to express your philosophical or scientific views. Just be respectful, open minded, and enlightened.  

1. Was Fischer correct or wrong about, "e4 is best by test."?
2. Where are the other 90 games?
3. Why was the match closed/disclosed as opposed to being public?
4. Would this same computer a.k.a. Alpha Zero (Or whatever it is) do better or worse with algorithm, databases, openings, endgames, etc and would the chess engines we have today do worse without those same tools?
5. How has this match affected your views on chess, computers, and the way you play as well as prepare for games or tournaments?
6. Is Chess a dead game?
7. Is Chess solved, at least theoretically/philosophically speaking (scientifically, the answer is still no)?
8. Is AlphaZero the "Carlsen" of computer chess in the relative dominance and approach that it has taken?
9. Should humans take this approach relying on intuitive analyzing instead of opening prep, calculation, etc?
10. As a result of this match, is the future of chess good or bad? Why? And feel free to explain your reasoning. 

Have fun discussing, but remember to be respectful of the opinions and views of others. [Comments will be monitored and disrespect, trolls, spam, etc will be blocked. You are warned.]

???

1. Fischer was neither wrong nor correct. Human players play openings according to their style/what positions they feel more comfortable with. You might evaluate it from a statistical point of view, but the differences between e4/d4 openings are very tiny. AlphaZero preferred to play d4 after some time and c4 too, even though it played more e4 inbetween and other openings too. There is no answer if it still would play them after some more training sessions or not.

2. The Deep Mind department only published a sample 10 games in their paper, the rest is somewhere in their databases most likely.

3. Only the Deep Mind team knows why. We also don´t really know how long they worked to optimize their neural network algorithms for chess.

4. AlphaZero is the neural network algorithm (a program) that runs on supercomputers that use TPUs instead of common CPUs. Those TPUs are not available to the public and are more powerful and efficient than common CPUs. AlphaZero doesn´t need an opening book or endgame tablebase unlike the programs like Stockfish to play better at those stages of the game. It gathers its "knowledge" by millions of training games in self play, you need to look up how they work.

5. It hasn´t much, we now understand that using neural networks is a feasible and probably better approach for chess programs than the common alpha/beta algorithms, even though the computing power is still huge for it to aquire such a high level of play.

6. No, people still play chess. And they will unless humans can learn playing perfect or near perfect chess.

7. As you said, no. But we know for a long time already that the tendencies are towards a draw with perfect play, not just after this AlphaZero vs Stockfish match.

8. Don´t really get what you mean, even though AlphaZero might be the strongest known chess entity as of today I don´t see any parallels to Carlsen, or his approach whatever it might be.

9. Maybe you misunderstood, but AlphaZero still calculates and it does it a lot. The 80,000 positions per second refer to the iterations it is able to process on that powerful hardware in one second. Given that Stockfish probably would calculate millions of positions on same hardware you can conclude that the AlphaZero algorithm is much more complex and requires more computation time per position. What use can humans make of it? If AlphaZero can reduce or neglect the weaknesses that normal programs like Stockfish have in some positions it will become a more powerful tool in their home preparation, nothing more. 

10. As said as long as humans still make mistakes it shouldn´t be a problem. Although possible problems arise if such powerful programs which also have human like play will become available to the public, the detection of cheaters might become even harder than it already is.

"8. Don´t really get what you mean, even though AlphaZero might be the strongest known chess entity as of today I don´t see any parallels to Carlsen, or his approach whatever it might be."

I apologize if I was not clear on the question. What I meant by that question was the fact that Carlsen puts little if any emphasis at all on the "Openings" aspect of the game. He prefers just to play as best as possible. In some similar manner, AlphaZero taught itself in 4 hours and comepletely outclassed the a chess engine with a 3400 rating (more or less) without the use of databases, opening books, engame table-bases etc. Other than that, I doubt there is any other parrallel between the two. But to be honest, I think humans should rely on intuitive analysis more than calculation because apparently, the results are better for humans and computers.

9. Maybe you misunderstood, but AlphaZero still calculates and it does it a lot. The 80,000 positions per second refer to the iterations it is able to process on that powerful hardware in one second. Given that Stockfish probably would calculate millions of positions on same hardware you can conclude that the AlphaZero algorithm is much more complex and requires more computation time per position. What use can humans make of it? If AlphaZero can reduce or neglect the weaknesses that normal programs like Stockfish have in some positions it will become a more powerful tool in their home preparation, nothing more. 

Interesting, the three articles I read all seemed to indicate no algorithm was involved with AlphaZero. In fact most of them stated that the algorithm was based on the specific mathematical programming to calculate and analyze millions of positions. This was in contrast to AlphaZero, which only analyzed around 80,000 positions but if I am not mistaken (which I probably am), Alphazero did not function with the use of algorithm. Or I think a better way to say this is that Alphazero had its own set of algorithms in contrast to other chess engines which use algorithms created by humans. Also, this is a question. Did Stockfish play without the opening book and endgame table-base and how do you believe it would have performed with it, better, the same, or worse?

I would have been more impressed if the test had been conducted on a level playing field.

When one of the competing programs is running on a super-computer and the other one is running on a machine less capable than my home computer, a victory for the super-computer is not an Earth-shaking surprise.

Interesting perspective. Thank you for sharing. 

Perhaps I should qualify that statement by pointing out that AlphaZero would probably have won a fair competition, too... it's the conditions of the match that disappoint me, not the AI program. I suspect that Google's PR department was more concerned with ensuring an overwhelming victory than with gathering objective data.

I think there is some accuracy to what you said. By the way, do you have access to their published research paper or link?

I wonder how much  the supercomputer time for the 100 games cost.

 It's Google, I imagine they just bought a computer outright. 

8. Carlsen puts a lot of emphasis on opening play as well as other phases of the game too. He cannot allow himself to be left behind other top players. You probably didn´t see his World Championship matches where he often nullified any of his opponents opening preperation without much effort. If you want to play as best as possible you need to be a beast in any game phase. Characteristic for Carlsen is his superior endgame play which doesn´t mean that he is not so good in any other game phase. Quite the opposite.

9. The paper is here and of course it uses an algorithm: https://arxiv.org/pdf/1712.01815.pdf

And yes Stockfish with an opening book and endgame tablebase play stronger than without. It needs to be said that the parameters they chose for their match were also dubious. A strange time control with 1 minute per move and somewhat low hashtable size (1gb) for the 64 threads. Also they played on different hardware, so the conditions weren´t equal.

The arxiv.org paper?

https://arxiv.org/pdf/1712.01815.pdf

I´ve never studied openings. Glad to see the most powerful player ever demonstrated  I was right: just do the best move from the very beggining of the game 

6.  Is Chess a dead game?

See this post:  https://www.chess.com/forum/view/chess-buzz/alphazero-will-people-treat-chess-the-way-they-treat-tic-tac-toe

 Gentleman, there is no longer a match computers vs. humans. Back in the 90's there was some compettition (AEGON and things like that). Nowadays, a decent smartphone outclasses Carlsen, let alone normal human beings. Anand stated that - 10 years ago - he could argue with his programs and sometimes make his point, but nowadays it's silly. You just have to understand why it says a certain move is better. Resistance is futile. This said, the A0 paper  was a deception, it's not clear how strong the hardware was in comparision to Stockfish set-up, what the disadvantage was for SF without his openingbook, A0 doesn't use one either, but it's OPTIMSED for it. It's not even clear whether Google took on SF, knowing that it had certain flaws which would lead to this result. Anyway, the outcome was clear. I propose the following  let's have a match SF - A0 where both have their best setup possible. SF running on a supercomputer (Sunway TaihuLight or something that size) with openingbook and everything you can imagine. A0 running on whatever Google has in store ( you can be sure they have some impressive hardware too) and see the result. If A0 still wins, we can only conclude that their approach is  better.

Better they collect game data from corespondence chess rather than from live chess.The game quality is better in correspondence because the players have maximum time in thinking.Also we rarely got draws game in correspondence.

Technically, although openings are not "studied" by the engine, it still prefers certain positions such as Queen's Indian over King's Indian or Ruy Lopez Berlin over Sicilian Najdorf. But, chess openings that are relatively equal 0.50 (Half a pawn for white) to -0.50 (Half a pawn for black) will still be completely playable. 

True, do you have any questions or personal opinions.