Engine styles for Analysis

[guyhayton]

I have read on various forums that engines have different styles for analysis purposes... aggressive, positional play, draw-ish etc. etc.
So effectively people have a mix of chess engines to represent "human personalities" and will mix engine choice depending upon either the desired outcome or to compare viewpoints.

Whilst a novice, I wish to get more involved in analysing my games and will practice using Aquarium/Chess Assistant on GM and IM games.
Too much choice would be natural thought - that is download and but all available engines. But I hoped to focus and get some experienced advice from others.

I currently have:

Houdini 2.0b
Deep Rybka 4.1
Deep Shredder 12
Critter 1.2
Komodo 3

What personalities would people place against these five? Is there enough variety? I don't mind spending money as long as I know it will be of benefit.
Any direction would be appreciated...

Cheers
Guy

[Jeremy Bernstein]

I have read on various forums that engines have different styles for analysis purposes... aggressive, positional play, draw-ish etc. etc.
So effectively people have a mix of chess engines to represent "human personalities" and will mix engine choice depending upon either the desired outcome or to compare viewpoints.

Whilst a novice, I wish to get more involved in analysing my games and will practice using Aquarium/Chess Assistant on GM and IM games.
Too much choice would be natural thought - that is download and but all available engines. But I hoped to focus and get some experienced advice from others.

I currently have:

Houdini 2.0b
Deep Rybka 4.1
Deep Shredder 12
Critter 1.2
Komodo 3

What personalities would people place against these five? Is there enough variety? I don't mind spending money as long as I know it will be of benefit.
Any direction would be appreciated...

Cheers
Guy

My opinion: any of these engines will kick you around the room tactically, whatever their position in a list of tactical ability is. Any tactic that a human can reasonably find, any one of these engines will see easily. In terms of positional evaluation, Komodo is, IMO, leaps and bounds above the others. I base this on the fact that Komodo's evaluation of positions in my own games most closely matches the evaluations of my (human) chess coach. That is, when I can't go over a game with my coach, Komodo generally serves as a decent substitute, if there's something I didn't understand.

If you are analyzing human games, and not playing matches against other engines, and not playing computer-assisted games against other computers/computer-assisted players, I would (personally) use Komodo + Critter (and Critter only as a sanity check). If you're doing those things, Uly would be the better guy to comment, but you can read his recent thoughts here (read the whole thread starting at this node).

I personally think that software like Aquarium encourages you to collect engines, set up a bunch of engine-specific trees and spend lots of time generating, guiding, and maintaining/pruning analysis, rather than simply finding a solid engine or two, analyzing, understanding and then going off to play more games. If you have a spare computer with lots of unused cores, maybe the typical Aquarium methodology makes sense, too, but I've never managed to make it work for me on a good laptop.

Jeremy

[Uly]

If you are analyzing human games, and not playing matches against other engines, and not playing computer-assisted games against other computers/computer-assisted players, I would (personally) use Komodo + Critter

Indeed, for such usage the strength of the engine doesn't matter, as long as it's strong to beat the opponents you're facing or the players of the game analyzed. For analyzing my own games, the SMIRF engine sufficed.

Here's a common scenario:

Say, you get some spectacular advantage as white against some opponent but your game ended in a draw, so you go offline to check the game with your engines. All of them point out the blunder, you gave a check to the opponent king to try to win a tempo and missed a variation that your opponent didn't, he played it and saved the game.

After the blunder the position was 0.00. After, c7, threatening promotion, the eval is more than 2.00. All the top engines and their dogs suggest c7 as the best move, and maybe, objectively, it is, but it requires a lot of precision from white, and the human doesn't really know how to play the positions, and if he played c7, later on he would have blundered and the game would have been drawn anyway.

Now, what's happening on this position is that a black Knight was attacking a Rook, and the player thought he could save it after the check but was wrong.

Say, the Pro Deo engine suggests that instead of c7, the player just saves the rook, and delays c7 for one move, and the human finds the variations very easy to follow, and Rb1 to be the clear best move to play, because the win is simplified. All the top engines complain and say Rb1 is only 1.20 (a 80cp blunder), but both moves are winning and the player is better going with the one he understands. For this, Pro Deo outmatched all the top engines, regardless of ELO or if with c7 the other engines would have beaten it in fewer moves, that doesn't matter.

In positions with similar patterns the player is going to save the rook instead of giving useless checks that give away the win.

tldr version of the example: When there are many moves that are best on a position, and all engines are favoring one, it doesn't mean it's the best, the best is the one you like the most and can understand, that leads to positions you have ease to find the best moves.

My point is that for analysis of own games I urge people to try several engines and stick with the one that gives the moves that are easier to understand to you. If Critter or Komodo provide such moves, that's great, but I've found engines of very high elo tend to pick some kinds of moves that sometimes seem to make no sense aesthetically and that don't seem to provide patters easily to remember once you're on real games, among the zillions of engines available there's probably one that picks your moves and the positions you like to play most of the time, and it's worth finding it, it doesn't matter if its elo is 2000 or less on the rating lists.

(For the actual question of the OP I replied on the sister thread at RF).

[Jeremy Bernstein]

If you are analyzing human games, and not playing matches against other engines, and not playing computer-assisted games against other computers/computer-assisted players, I would (personally) use Komodo + Critter

Indeed, for such usage the strength of the engine doesn't matter, as long as it's strong to beat the opponents you're facing or the players of the game analyzed. For analyzing my own games, the SMIRF engine sufficed.

My point is that for analysis of own games I urge people to try several engines and stick with the one that gives the moves that are easier to understand to you. If Critter or Komodo provide such moves, that's great, but I've found engines of very high elo tend to pick some kinds of moves that sometimes seem to make no sense aesthetically and that don't seem to provide patters easily to remember once you're on real games, among the zillions of engines available there's probably one that picks your moves and the positions you like to play most of the time, and it's worth finding it, it doesn't matter if its elo is 2000 or less on the rating lists.

In fact, this is why I don't recommend Houdini or Rybka or even Critter as a primary analysis engine -- they're a) too strong and b) too highly tuned toward computer/computer tactical warfare. They are too computery, for lack of a better word, and they will make stupid suggestions that gain you 3/10th of a pawn after 25 plies of perfect play from both sides, whereas your move was actually perfectly solid and totally human.

Engines which provide comprehensible information about real chess are Komodo, HIARCS, (apparently) Junior (don't have it), and certainly many others (I'd give Glass a try, as well). Just because Anand has a high Elo rating, doesn't mean that he has nothing to teach you about chess (apropos Larry Kaufman's recent comments about Komodo vis a vis Anand) !

The moral is: don't spend a lot of money on engines; do take some time to analyze your games with the engines you have and critically examine their output -- is it useful and clarifying? or is it simply detailed and confusing?

Again, any engine you'd consider is going to find tactical blunders (which will be the majority of your errors). An engine like Komodo (or HIARCS) will, from my experience, also help you understand where you went wrong with your positional gameplay, and this has been very valuable.

jb

[Uly]

If Komodo has managed to achieve that while keeping good results against top engines, that's very impressive. But I think Komodo would play c7 in my example, just because it's the fastest win and I'd doubt Komodo would miss it.

This is just a side effect to gaining elo, in real human games, swindling is sometimes very important, and top engines aren't going to swindle, if they see a refutation, they assume the opponent would see it as well and don't go for the trap, even if the refutation only loses 0.30 centipawns, no matter if the engine is Hiarcs or Junior. That's why I find it worthwhile to find an engine weak enough to not see the refutation (like the human won't), and that's got to be one with lower elo than the top engines that have been mentioned. Probably an older version of Hiarcs or Junior are going to be better for this than latter versions, at least, Hiarcs Paderborn 2007 was much better for this than Hiarcs 12.1, and I wonder if Hiarcs 13 has gotten its mojo back.

(I see Hiarcs has a Swindle mode, in my experience it doesn't do what you'd expect it to do, like Hiarcs's "learning")

[Jeremy Bernstein]

If Komodo has managed to achieve that while keeping good results against top engines, that's very impressive. But I think Komodo would play c7 in my example, just because it's the fastest win and I'd doubt Komodo would miss it.

This is just a side effect to gaining elo, in real human games, swindling is sometimes very important, and top engines aren't going to swindle, if they see a refutation, they assume the opponent would see it as well and don't go for the trap, even if the refutation only loses 0.30 centipawns, no matter if the engine is Hiarcs or Junior. That's why I find it worthwhile to find an engine weak enough to not see the refutation (like the human won't), and that's got to be one with lower elo than the top engines that have been mentioned. Probably an older version of Hiarcs or Junior are going to be better for this than latter versions, at least, Hiarcs Paderborn 2007 was much better for this than Hiarcs 12.1, and I wonder if Hiarcs 13 has gotten its mojo back.

(I see Hiarcs has a Swindle mode, in my experience it doesn't do what you'd expect it to do, like Hiarcs's "learning")

Any decent chess coach will tell you that swindles and refutable traps are nothing but a cheap hope that your opponent is going to miss something -- if you play like that consistently, you'll never get past a certain level. Any decent tactics book will exclude such "traps", and game collections generally include an annotation to the effect that the player just got lucky. I'm afraid that I can't follow you down that road. I would bitbucket any engine that gave me refutable moves as "good" -- we're supposed to be learning something here, right?

If you're talking about "practical chances in an uncertain or bad position", I think that's something else, and in fact something that engines have a hard time with, because they have a psychological component. But if we're talking about "equal position, let's play a trick and win some material," I can't agree.

jb

[AnthonyTheSage]

The main difference between engine play and human play is that engines have no fear. They play extremely sharp lines without even a second thought. Humans ( most humans anyways) tend to stick to more positional long term plans. I think this is the only real advantage that humans have left, we can generalize a position very quickly. Which actually allows us to see further ahead in certain position. Although not exact, we can generalize the outcome.
For giving advice on solid moves I would recommend Shredder 12. I would say its the most positional and reasonable out of the bunch. The styles between the rest seem to have become very similar. I think since the ippolit etc. source codes have been floating around all the top engines seem to have lost their distinct styles. But for finding refutations I don't think you can beat houdini 2. It's a tactictal monster, and if theres something to be found in a tactical position houdini will usually find it.

[Uly]

Any decent chess coach will tell you that swindles and refutable traps are nothing but a cheap hope that your opponent is going to miss something

Yeah? What about missing a move that only top engines would find and that even Anand is going to miss, due to its huge depth?

Would you play the move that scores 0.60 with the engines instead of the 0.30 one, even if the engines show the best the opponent can do is play Be6 in front of the d pawn, something the human is probably not going to do because of the c Bishop getting locked down, just because there's some deep tactic the engines show in where objectively, Be6 is the best move?

I claim that engines can get too strong and start self-refuting their own ideas, and showing refutations that no human would ever find on the board.

The best moves to play against other engines are not always the best to play against humans, that's what swindling is about. That's why you can play dangerous gambits against a human and not against an engine. The engine is going to find all the refutations and kill you. If the human hasn't played the positions before, you're going to beat him with your knowledge of the gambit alone.

"equal position, let's play a trick and win some material," I can't agree.

But if we're talking about some super-dangerous move in where the engine struggles to see the refutation, and most humans can't even calculate that deep and see it coming, then that's probably the best move against a human. How many of Tal's sacrifices were sound? I recall examining some of them and finding that they weren't sound, but they were winning him his games, playing such tricks isn't "hoping that the opponent misses it", it's knowing that no human would manage to find the refutation on the board.

Weak engines have less problems with self-refutations on their ideas.

[Jeremy Bernstein]

Any decent chess coach will tell you that swindles and refutable traps are nothing but a cheap hope that your opponent is going to miss something

Yeah? What about missing a move that only top engines would find and that even Anand is going to miss, due to its huge depth?

Would you play the move that scores 0.60 with the engines instead of the 0.30 one, even if the engines show the best the opponent can do is play Be6 in front of the d pawn, something the human is probably not going to do because of the c Bishop getting locked down, just because there's some deep tactic the engines show in where objectively, Be6 is the best move?

I claim that engines can get too strong and start self-refuting their own ideas, and showing refutations that no human would ever find on the board.

The best moves to play against other engines are not always the best to play against humans, that's what swindling is about. That's why you can play dangerous gambits against a human and not against an engine. The engine is going to find all the refutations and kill you. If the human hasn't played the positions before, you're going to beat him with your knowledge of the gambit alone.

"equal position, let's play a trick and win some material," I can't agree.

But if we're talking about some super-dangerous move in where the engine struggles to see the refutation, and most humans can't even calculate that deep and see it coming, then that's probably the best move against a human. How many of Tal's sacrifices were sound? I recall examining some of them and finding that they weren't sound, but they were winning him his games, playing such tricks isn't "hoping that the opponent misses it", it's knowing that no human would manage to find the refutation on the board.

Weak engines have less problems with self-refutations on their ideas.

You're only winning because your opponents are using their computers less than you are. How can you be proud of that win?

As you mention, a lot of Tal's sacrifices are unsound and worked due to the psychological effects. It didn't take long for many of his opponents, particularly Botvinnik, to adapt, though, did it? Unsound is unsound. I understand your point of view, I just disagree -- it insults the intelligence of your opponents. In your case, it also underestimates their ability to use a computer. At some point, you'll hit a big, hard wall with that attitude, IMO. But if you're not there yet, I guess there's no harm in playing like that.

Jeremy

[Uly]

You're only winning because your opponents are using their computers less than you are. How can you be proud of that win?

How do you know that my opponents use their computer less for their analysis of postmortem games?

Engines only can show me patterns, but if I learn them, on the OTB games, it is me who is playing the moves. For those games, then engines are switched off and are unrelated.

Your statement makes no sense, that's like saying "You're only winning because your coach is training you better than the coach of your opponent trains him. How can you be proud of that win?"

I understand your point of view, I just disagree -- it insults the intelligence of your opponents. In your case, it also underestimates their ability to use a computer.

No, I don't think you understand. Even, I'm not sure we're talking about the same things.

This is what I'm talking about:

A human plays a human, unassisted. There's some game result.

The first person takes the game and checks it with a computer at home, sees what ways it could have played better, or in what kind of patterns he was missing a beneficial theme, and improves. Maybe he even decides to ditch the opening played and go for something else.

The next day, the humans play again, and chess is so complex that it's very unlikely the patterns appear again and the player can take advantage of what he learned with the engine, but he'll possibly play better, and at some future game, the motif will appear and the player will get better results than if he just played but never analyzed his games.

My claim is that weak engines can be better than top engines for this process if they have a style that fits the human. Furthermore, if the user sticks with Komodo, whatever style Komodo has, it's going to disrupt the style of the human. Say for example, that Komodo likes to push c4 in some kind of positions before playing Nc3, even though Nc3 is a perfectly valid move and the one the human liked, but eventually the user sees it so much that he starts playing c4 instead of Nc3 in his games. That's style disruption, and it happened to me with Pro Deo, that would make moves that would stop the opponent from pinning my knight, and would play knight to the border of the board to attack bishops. I thought they were the best moves, only later on I learned they weren't necessary, that I could play different moves, but my playing style was already disrupted, that's why I think it's best to look for the engine that agrees with your moves, and suggests improvement along the same lines, regardless of ELO, instead of a top one that will suggest totally different lines from what you play.

You make it sounds as if computers could be used while the game was going on, or something in where using more computers is going to give the human more wins. At some point this became the conversation of the deaf.