How good do chess engines play chess?

You answer the question 'do chess engines play better chess than humans?' That is not the question of this thread.

Speaking about chess AI, here is an interesting development. A neural network that learns to play chess more like humans, instead of relying on brute force.

Once one gets to the paragraph thas says "But while computers have become faster, the way chess engines work has not changed. Their power relies on brute force, the process of searching through all possible future moves to find the best next one." an immediate conclusion is that whoever wrote that piece has completely no clue about how engines work.

A subsequent claim that "out of the box, the new machine plays at the same level as the best conventional chess engines [...] On a human level, it is equivalent to FIDE International Master" confirms that conclusion.

There may be something useful in the reference: http://arxiv.org/abs/1509.01549

I just came across this thread, but I had been wondering this very same thing. Additionally, as humans can't play simply based on calcutaions but need to employ positional ideas that are born from a deep understanding of the game, do humans have the capacity to find or to develop more powerful, maybe even more fundamental ideas than now exist? 

Human and chess engine is much stronger than chess engine alone.

Also, sometimes humans solve problems which the best chess engines cannot solve.

putkey

@kkl10: not using brute force, but only a dataset of 175 million positions? The best chess engines play at the level of the super GMs? Not a very precise article to say the least. But a neural network can be interesting in order to learn better how we humans think. It might gap the bridge between binary silicon power and our very fine tuned neural networks.

@batgirl: that is indeed the source of my question. Chess engines make better use of the ideas humans have put in them than humans are capable of doing themselves. That implies that there is still a lot of room for learning improvement in chess - and hence in other fields of knowledge as well.

..."super GMs" is a figure of speech, obviously. Not difficult to interpret what it means... are you guys getting the point of the article?

The neural network was pre-given a library of positions, and then left to play against itself for a certain period of time. During this time, it learned to recognize the best positions. What's remarkable here is that it developed its own heuristics. Which is to say that after a learning period it doesn't have to rely on pure calculation like other chess engines; it relies on pattern recognition of the positions, a bit more like humans.

This is well-explained in the article. I found the comments interesting too.

Yes, I got the point of the article.

Neural networks are well known to have that capacity of self learning. There were already computers with primitive neural networks in the 60's. I learned about it in the 80's during my study psychology. They had a research program developing neural networks.

I don't know if chess engines can learn from their own mistakes, but I guess that they can evaluate the same position differently based on newly collected information from their own games. If that is called learning, then I say 'Yes, they can.'

This article might be of your interest: http://chess.stackexchange.com/questions/1003/a-purely-self-trained-chess-ai

or this one:

https://chessprogramming.wikispaces.com/Learning

Take a look at the end position of Peter Svidler en Wei Yi:

Stockfish is playing this endgame in a way I would never think of. Why is it better than the way I see? My idea is to keep the king and rook as long as possible to the back rank and to bring white in zugzwang. 

And if you take a look at the final position of Eljanov - Karjakin:

Stockfish wants black to play Kc7 after Ke4. But it is a forced draw in 2 moves, because Karjakin has created a position in which the capture of the pawn is inevitable. Stockfish does not seem to understand it. Very strange imo, because it is a basic technique to make a draw.

After Ke4, there is just one losing move for black, so the engine picks any of the ones that draw. It has no reason to prefer any of them.

1...Kd8: Draw
1...Kb8: Draw
1...Kc7: Draw
1...Kb7: Draw
1...Kd7: Draw
1...Nb2: Draw
1...Nc3+: Draw
1...Nf2+: Draw

1...Ne3: White wins in 14 moves

The tendency for programs to have no preference for different moves if they are certain to lead to the same result can lead to some quite funny situations. A situation once described by Tim Krabbé went as follows if I can recall it correctly. Two programs were playing each other in a tournament, and one of them had K+R+R+P vs K+R, obviously a trivially winning position. Both programs were also equipped with a tablebase for the endgame K+R+P vs K+R which they would therefore be able to play perfectly. The programs then proceeded to play in the following ridiculous style: the stronger side kept attempting to sacrifice one of its rooks to end up in a K+R+P vs K+R position which would be winning according to the tablebase, and the defending side kept declining the rook because it too saw from the tablebase that the resulting position would be lost.

@Brian-E -- indeed, a hilarious effect. Reminds me of many jokes about mathematicians reducing a problem to one with a known solution, e.g.

A mathematician and an engineer are on desert island. They find two palm trees with one coconut each. The engineer climbs up one tree, gets the coconut, eats. The mathematician climbs up the other tree, gets the coconut, climbs the other tree and puts it there. "Now we've reduced it to a problem we know how to solve."  

 made my day :)

awesome indeed :-)

I saw several positions which are wrongly evaluated by stocckfish.

It is abosultely a draw position but all engines show white is going to win and white's score is around 6.50.

see Another position

Stockfish says, it is a draw position, actually it is winning position for white. Even after going to 25 depth , it shows 0.00 score. After making Re1-e5 , stockfish gets understood that it is a decisive position. Funny fact is that even Rybka can identify the winning position for white.

@LegoPirateSenior: after Nf2+ you force a draw within two moves, where all other moves give white still a way to win. It is like choosing between in mate in two or a huge winning advantage and telling they are the same.

@Earth64: After Re5 I would play dxe5 as black and expecting to win. After Rf4 could I not see a direct win either. Nor after Rg1 or Rf2 Qd4. Could you show us how white would win?

@LoekBergman&LPS:

LB:"where all other moves give white still a way to win" LPS:"there is just one losing move for black"

There is a serious disagreement here 

@mcris: lol, to complicate the matters even more: LPS and I both agree that the position is a dead draw and we will both agree with the validity of the mutual exclusive statements made. :-)

As long as there is a pawn on the board, white has a chance to win. If you will look at all variations of each move, then will you see pathways in which white will win. You will not see them after Nxg4.There is simply no mate possible with those four or three pieces (after white captures the black knight).

I would say that the best road for display how to get the draw is the way which is most forcing a draw.