Computers and endgames

Coming back to chess in recent months after a decades long sabbatical, I have been interested to see to what extent computers and engines have developed. Clearly the very strongest of these have come leaps and bounds since the days of Deep Blue, the appearance of which caused quite a stir at a time when computers out there on the market couldn't even beat stronger club players, never mind Masters. In those days it was acknowledged (then as now) that computers had great tactical 'ability'. They were, however, weak in endgames where to humans the trees of candidate lines tended to be rather sparser and continuations more obvious. Computers, it was thought, were particularly weak in this phase of the game because they continued to look over all variations (rather than subjecting them to the more rigorous forms of triage exercised by humans at this stage) and had little strategic  (as opposed to tactical) 'understanding' (very important in the endgame).

Having read about developments to computers and engines, I thought that this might not be the case now, not just for the top level computers and engines but also for the ones more popularly available. I was surprised, however, this morning to discover just how awful the chess.com Computer Level 7 was at endgames. Level 7 played a cogent, positional middle game against me in which I had to free myself from near zugzwang, sustaining what momentum I was able to cook up with a pawn sacrifice that left me in an endgame a pawn down but where I was compensated to some extent by the fact that we had opposite color square bishops in what looked to me like a cast iron draw. Interestingly 'she'(?) declined two offers of a draw from me (made in the form of offers of a third position repetition). A pawn up and with opposite color bishops as insurance, Computer Level 7 still managed to balls up the endgame giving me in  the end an easy and not very satisfying win.

Why is it, I ask myself (and you), that computers (not those at the very highest level) are still so a%$* at endgames? I would have thought that this at least would have been addressed in the last three decades. I will post the game below in case anyone is interested.

Excellent post and question.  I would say because it's not so easy for the computer to determine which paths to examine in greater depth.

I've seen engines make obvious mistakes before.  Maybe they weren't at full strength but complex endgames are common.

There are complete six-piece and partial seven-piece tablebases that engines can use in the endgame.

You make a distinction between "outplaying" and "out-calculating".  But isn't out-calculating a valid way of outplaying someone?  Human players can out-calculate other humans.

I can imagine a computer saying (of a strategic position that it is losing) "It looks like humans are outplaying computers but they are just out-intuiting them".

An engine has a decently strong but not unbeatable positional evaluation, combined with a 4000-strength tactical monster.  As a result, while a human can never match the engine's tactical ability, a strong human player may have an advantage in certain very quiet positions and certain endgames.

Interesting responses everyone. I think JMurakami's point about the evaluation of long term threats is particularly salient and I was thinking the same myself. I guess I was really expressing surprise that after thirty odd years (during which time my engagement with chess was minimal) any developments designed to address this particular shortcoming (if there were any significant developments) hadn't trickled down to the likes of Computer-Level 7.

Part of the issue is the question of how to make a computer play at a human level.  One obvious answer is to reduce the analysis depth - thus bringing back the problem with endgame weakness.

I think the programmers have discovered that only reducing the analysis depth is not a very good method of simulating a human level of play. A more natural method appears to be to limit the number of analysis nodes/sec, coupled with the addition of a randomness factor in the evaluation term.

Yes, reducing the analysis depth is only part of the solution, but it is a legitimate part.  Limiting nodes/sec accomplishes more or less the same thing.

The randomness factor is also a common technique.  This won't result in major tactical blunders, but it can result in positional errors which can be deadly in an endgame against a decent player.

Another interesting method I once saw was to randomly remove some branches in the search tree; a small number at low levels, and a much larger number deeper in.  That was intended to simulate human-style tactical oversights at deeper levels.  However, I think it would still tend to result in typical weakened engine play of a series of ultra-strong moves followed by a blunder.

You need to be a freaking god to even notice engine's weaknesses in the endgame, appart from thinking a wrong colored bishop + pawn is winning.

I beg to differ on that, BP. If you go through my game and consider some of Level 7's choices in the endgame I think you'll be surprised how poor (by the standards of a human playing at Level 7) they actually were.

vickalan - I don't think I'm particularly good at endgames. As I said in my reponse to BronsteinPawn I felt that the engine's standard (if considered against a more or less comparable human player) very noticeably dropped.

Are you trolling me? Computer level 7 is programmed not to play its best. Im talking about Stockfish on his full power, not about a tweaked engine.