I don't know very much about openings or high-level play in chess. Did computer analysis ever rule out formerly popular moves as bad, or create new openings?

I heard yes.

I read this article a few days ago; it might have been linked from this very forum. I found it relevant to your question.

http://www.businessinsider.com/anand-on ... ss-2013-11

_________________
La victoire est un hasard, la défaite une nécessité.

It is pretty rare that a move formerly considered reasonable has been completely refuted by a computer and abandoned (I can't think of an example off the top of my head). Consider the Go analogy; AlphaGo probably isn't going to discover that some established joseki actually ends in disaster for one side due to some incredible tesuji.

What happens more often is that computers can pretty exhaustively search the possibilities in a position or line and confirm that it is promising or not. A typical opening line might obtain some static advantage for one side but concede some dynamic possibilities to the other, and the computer can check whether those dynamic possibilities really come to anything in the end with best play. If not, the player can go ahead and play the risky move (if he remembers or can reproduce the computer analysis over the board!). As a result, chess style has become more "concrete", focusing on whether moves work out in actual variations, rather than evaluating them with rules of thumb and pattern-matching. (Of course people have always done both; it's just that the balance has shifted more to the former.)

One interesting result is that many top players now play openings that are less sharp and concrete, avoiding computer-assisted preparation on the part of the opponent and trying to "just play chess", relying on their more human strategic skills. Magnus Carlsen (the current world champion) in particular is the poster child for this approach.

As I know that there are professional opening books available such as the Fritz Powerbooks that extend up to move 30, I would say yes. As I have recently replayed a commented game that started the comments with "Magnus Carlsen tries a novel approach in this opening; this move has not been played before" at move 27 (!), I would say yes.

IN the days that I played chess seriously (in the 90's), openings would be 10-12 moves long (12 moves on both sides), with some at 15-ish being very long ones (edit: at least for amateurs). Huge opening books and endgame tablebases were only available for people with a lot of money, and a lot of hard drive space.

Opening books and endgame tablebase are getting bigger and bigger. Opening books at 30 moves, tablebases at 7 pieces, as new storage methods are developed (Thompson, Nalimov, Syczygy, Lomonosov), makes the middle game almost a trivial matter for computers. It's possible to play an opening to the end, play a shortish middle game, and hop right into the endgame tablebase; that is the reason why some of the strongest engines on powerful hardware can now announce stuff like 'mate in 30' at move 30-35 or so. They are powerful enough to leave the opening book and see straight through the middle game right into the endgame tablebase, if there aren't too many pieces left after such a 30 move 'opening'.

The computer has not only changed chess openings, it has changed everything. To a computer, chess holds no more mysteries. If you want to know if your move was right, you ask the computer for a 5 minute analysis, and the answer will be definitive.

This is really interesting! :-)

Chessbase Fritz Powerbook 2016

Over 21 million opening positions.

6 piece endgame tablebase, for free download
7 piece endgame tablebase for purchase or online.

Good luck playing chess against an almost perfect machine

I think this idea is subtle. Just because computers can beat us at chess (and probably in the near future go) doesn't mean they play really well in an absolute sense. It might be that we are just really bad ("puny humans!"). There still might be a long way to go in playing strength. For example in computer chess the top program beats the 3rd best with some regularity and ratings of computers chess machines are going up rapidly.
see: https://praxtime.com/2014/03/24/chess-technology-progress/

For example, suppose that deep learning ala alphago get to the point that they can be pros every time. It might happen that some completely different approach might crush alphago and that approach could play go very differently than humans do, while alphago plays a lot like humans since its neural net was trained on human play.

That engines are stronger than humans is clear. What I always would have liked to know is how much stronger the engines actually are. Because humans and engines don't play in the same tournaments, the ELO scales are not comparable. Also, you can't just have Magnus Carlsen play Stockfish or Komodo in a 6 game match; he'll probably lose by 6-0. Because of that, you can only say that the engine is stronger than Magnus Carlsen, but not by how much; 300 ELO? 500? Even more?

What we would need to do is set up a huge tournament. Let's say, the top 16 players, and 16 engines, taking the top 4 engines, and then 12 that become progressively weaker. Then set up a 20 round Swiss tournament, just as if all participants are humans. So, sometimes humans will also play humans, and engines will play engines.

Against the weaker ones, some of the strongest humans will draw, or maybe win. Those weaker engines however, will also draw or sometimes win against somewhat stronger engines; and that pattern will run all the way to the top engine. Thus, even the top engine will probably not have a 100% score.

As the engines and humans are playing in the same tournament, it's now possible to calculate the ratings of the engines; the one that scores the same number of points as Carlsen has a performance rating of 2851. For an engine, performance rating == rating, because an engine doesn't fluctuate up or down. (Not by a meaningful margin anyway.)

Now you would have the 'real' ELO-rating of at least 16 engines, and thus you can re-calibrate an authoritative list such as the CCRL using those ratings as calibration points. Then you would know exactly where humans top out, what engines are stronger than humans, and by how much.



It's certainly possible that new methods of playing Chess and Go will be developed that are even better than what we have now. Take a look at Stockfish for example. Note that Stockfish 7 64-bit is 37 ELO points stronger than Stockfish DD 64-bit 4CPU. What this says is that version 7 is stronger than version DD while only using 25% computing power.

This can only mean that version 7 has had great algorithmic improvements compared to version DD and older. Who, 20 years ago, would have thought that chess engines would become so advanced and powerful that they can defeat anybody except maybe the strongest grandmasters when running on a budget phone?

The same will happen with Go software now that a good, workable way of playing Go by a computer has been developed. Now it still needs something like an AlphaGo cluster to beat Lee Sedol, just as IBM needed the Deep Blue dedicated chess supercomputer to beat Kasparov. Go programming now is where Chess programming was in 1997, but the software is advancing much faster than back then. In 10-15 years from now, the top Go grandmasters will be having a hard time against a budget smartphone as well.

Could Magnus Carlsen defeat a computer with a handicap? Say the computer was down a pawn. A knight?

I heard chess programs are already at least 1 pawn handicap past the best humans.

Probably.

It has been tried against Rybka 3, who seems to lose a bit less than 650 points when always playing black and handicapped by removing the F7 pawn. Rybka, estimated rating 3200, achieved 3 draws, one win against Roman Dzindzi, rated 2550, with this handicap.

Assuming that a newer engine such as Stockfish or Komodo (rated at +/- 3350) would also lose 650 points when playing black and handicapped by removing the F7 pawn, the engine's rating would be 2700. Carlsen is rated around 2850. This is a table with winning probabilities:



So, Carlsen could possibly win a 10 game match by 7-3.

Milov, 2705, against Rybka 3, 4.5-3.5 win (several different handicaps)
Handicap match against Komodo 9, which is a much newer engine than Rybka 3, human still wins (and isn't even a top grandmaster).

The humans can draw or win when playing carefully, as soon as the engine is handicapped. It seems the estimate of an engine losing around 650 ELO with a one pawn handicap is in the ballpark.

The most recent handicap games against a top-level player were against Hikaru Nakamura (currently world #6) earlier this year.

The handicaps were pawn-and-move (Komodo played Black and without one pawn), pawn (Komodo played White and without one pawn), exchange (Komodo played White without a rook, Nakamura played Black without a knight), and 4-move (Nakamura played White and played 4 moves in a row before Komodo started to play).

The first three games were drawn and Komodo won the fourth. All games are at the link I provided earlier and are fun to play through if you are interested in top-level handicap chess.

My bold face. Even among human pro players a joseki might be abandoned because one side has what seems a very small advantage, e.g. one point or less. AlphaGo might abandon a joseki because of even smaller advantage, not due to the joseki being "crushed". However, can AlphaGo make such quantified judgments?

I can't speak about AlphaGo, but I would say yes; if not now, then it will in time. A chess engine is able to quantify positional differences into 1/100th of a pawn.

It's very normal that, even if material is exactly the same, a chess engine will denote an evaluation of +0.78, for example, i.e. the advantage is a bit more than having three quarters of an extra pawn on the board.

I've played games against Stockfish, set at 10-20% depending on the device I play on. Sometimes I lose a game without making any blatant mistakes; I 'just lose'. Then I let Stockfish on my computer analyze the game for an hour or so, and I'd see an evaluation histogram akin to this:

Move 10: -0.21
Move 20: -0.47
Move 30: -0.75
Move 40: -1.10 (while material is actually still even)

And then between move 40 and 50 you'd actually lose a pawn, the rating will fall to something like -2.00 or worse, and Stockfish will crush you in the endgame.

Wow... a 1-2 pawn handicap weakens Komodo by about 500 ELO. It's possible to draw, and maybe, win a game. However, it doesn't give a zilch about a 4 move disadvantage. It just keeps the position closed until all it's pieces are developed, and then starts to play a normal game... crushing you by an advantage that creeps up by 1/100th of a pawn for 30 moves. I wonder how many handicap moves a player like Nakamura would need to actually win such a game. In the last game, Nakamura actually is a paw UP, and still he's 6 pawns DOWN. That's a 7 pawn (more than a rook...) positional advantage for Komodo.

Clearly, there is a lot of improvement that can be made by humans when playing chess, but I wonder if anyone ever will have the capability to obtain an improvement that large.

Wow, that odds match between Komodo and Nakamura had gone completely unnoticed by me up until reading this thread. I've just browsed through the games. It's pretty much business as usual-as soon as the position opens up, the engine tears the human apart. It seems that a least knight odds are required, and even then I wouldn't put top money on the human.

As for the original question about openings having been put out of use by engine use. The question does not have a straightforward answer. No well established system has been refuted, for starters.

But there are a lot of subvariations that cannot be played anymore against an opponent who knows that he is likely to face it and thus is willing to prepare. This mostly applies to several cut-throat variations in the open Sicilian.

Apart from that, I have the impression that even the main line Najdorf Sicilian (https://en.m.wikipedia.org/wiki/Sicilian_Defence,_Najdorf_Variation) isn't being used much at the top level nowadays. Economical reasons are most likely an explanation - the work required to stay ahead of opponents is huge, not to speak of the risk of mixing up something and falling apart right away.

Since computer programs are significantly stronger than human players, the question as to whether an opening line has been "refuted" is a moot question. For humans playing against the top programs in chess, and soon in go, the human players will lose no matter what opening line they choose. Historically, refutation of opening lines was the same thing as saying that the opponent is guaranteed an advantage from that line. It might be correct to say that, for humans, chess has been refuted.

Chess players don't consider it moot. For an opening line to be "refuted" means that the opponent can achieve a clear advantage when both players are competent and of equal skill. This might be human against human or computer against computer. A line is not refuted just because a grandmaster could eventually win from that position against an amateur, or a computer against a grandmaster. It is still very relevant to human-human play when a computer finds a clear refutation to a line, since (good) players learn the refutation and the followup play, and it becomes very dangerous to play it in tournaments.