13

I have seen that engines aren't particularly strong in the openings without access to a book.

How strong are modern engines in the endgame when they are not allowed access to tablebases? E.g., can a computer typically win those difficult endgames such as KQ vs. KR and KBN vs. K in blitz games?

2
  • We don't really know what strong opening play without an opening book looks like. We have no basis for comparison since all the best human players study the standard openings. It would be an interesting experiment to take a couple of strong chess engines and constrain them both so that they are not allowed to reach any position found in say Modern Chess Openings after 8 half-moves and see what their play looks like. Maybe after you prune the tree of accumulated openings knowledge there aren't any good openings left.
    – Kyle Jones
    Oct 18, 2012 at 17:34
  • @KyleJones I think actually they test the chess engines strength by playing Chess960 (google it if unfamiliar), at least that was the case on the link above, I was disappointed because it took me a lot to download the PGN (60mb) and I was looking to improve my opening >_<
    – ajax333221
    Oct 19, 2012 at 23:04

3 Answers 3

9

First, I'm not convinced that engines are necessarily weak without their books. Without an opening book, modern engines might not play the most testing lines, or lines that would get them into a position where they thrive, but any top engine will still crush any human GM without an opening book.

As an interesting aside, if you leave your engine on for a while, it usually finds the Spanish Four Knights as the best line from the starting position, but I digress...

In the endgame, it totally depends on the engine and the position. Without tablebases, engines might not know that a rook pawn and wrong bishop is only a draw. If the programmers added that exception, then they'll figure it out, but otherwise, they might trade down thinking that they'll win with an extra bishop and pawn only to draw via the 50 move rule.

In more elementary endgames, like KBNvK, the stronger side will almost always win. The longest mating sequence is 33 moves if my memory serves. Most engines will stumble onto a forced mate in that time because their heuristics will have them make good moves that get closer to mate (and they can probably see 15-20 moves deep).

In terms of KQvKR, the stronger side will almost always win against a human. Humans are pretty bad at defending this position, and the computer will instantly see ways to pick off a loose rook. Against another computer, it's a little bit harder, but as Wes says in his answer, it is still likely that the stronger side would win (the Philidor maneuver is only ~5 moves long, plenty short enough for the engine to see it) given enough time.

7

Win against who?

I've seen KQ/KR draw between computers before, in 1 or 5 minutes (without egtb). They can drill pretty deep when there aren't many pieces on the board, but if the time control is fast enough, it's not always deep enough. Even GMs draw that one sometimes.

Still, I imagine a normal person might make a non-optimal move in such a time control, also, that could tip the advantage to the computer.

2

Tablebases add only a little elo to computers. The amount added is small enough that for many years Andscacs was one of the strongest engines around without endgame tablebases (it still doesn't have them implemented). Even for other engines with tablebase support implemented I'm pretty sure you can not provide it with an endgame tablebase, and they'll play almost as well. Sure, you lose some elo by not having tablebases, but not a lot (I don't know exactly how much, but I'd be surprised if it's more than 10 elo).

The brief answer to whether computers can win KQ vs. KR and KBN vs. K is yes. Consider, Stockfish on reasonably strong hardware will easily be operating at 10 million positions per second. How many KQ vs KR positions are there? Stockfish will search them all in a couple of seconds. There's no doubt that Stockfish without a tablebase will easily beat a tablebase at this endgame.

5
  • It is not just about number of positions but also about branching. Sometimes KBN vs K can take as many as 33 moves to win (with perfect play). And there can be as many as 29 moves available at each step. 29**33 = 1816075630094014572464024421543167816955354437789. Then "10 million positions per second" doesn't seem all that big..
    – wim
    Jan 21, 2020 at 5:49
  • @wim try it - take this FEN K1k1B3/8/8/8/8/8/7N/8 w - - 0 1 (which is 33 moves to mate) and put it into the computer at chess.com/analysis. You should immediately get a +49 eval. Try defending with black using the tablebases at syzygy-tables.info/?fen=K1k1B3/8/8/8/8/8/7N/8_w_-_-_0_1, see if the computer can checkmate you fast enough.
    – Allure
    Jan 21, 2020 at 5:56
  • I got 1. Ka7 Kd8 2. Bb5 Kc7 3. Nf3 Kd6 4. Kb6 Kd5 5. Bd3 Kd6 6. Be4 Kd7 7. Kc5 Ke6 8. Bb1 Kd7 9. Kd5 Ke7 10. Ne5 Kf6 11. Kd6 Kg7 12. Ke7 Kg8 13. Kf6 Kh8 and then the computer (which if I'm not mistaken is Stockfish 10; Stockfish 11 was released just yesterday) started reporting mate scores, no need to check further.
    – Allure
    Jan 21, 2020 at 6:04
  • Ok, but how do you know that chess.com/analysis is not using tablebases?
    – wim
    Jan 21, 2020 at 6:05
  • @wim if it were using tablebases, it would not report a +49 score - it would immediately give "mate in X moves".
    – Allure
    Jan 21, 2020 at 6:08

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.