![[nl]](/mwf/v12/flags/nl.png "Netherlands")
I wish to Appeal to your Expertise.
I found out this List:
http://www.chess.com/downloads/playing+programs+engines
I finally want to get my Games Analyzed ( in WinBoard? Or do you know a better free GUI for this Stuff? ) now that I Book Progress due to Chess Theory ( not Really in the Case of Openings
, I am sorry )
What ( or which of this List. It's Big though, you should not Search it for me) Engine that you know, Plays Passively, Defensively, Prophylacticial, Positionally, is not affraid of Draws ( But ofcourse prefers CheckMates huh ), and could be in this List?
Ofcourse, when there are Offensive and/or/nor Aggresive ( ( Exchange-)Sacrifices ) chances, one should use those Chances, but I am just not that of a Dynamic, Tactical Player - unlike most others -, Basically, and I don't want my Personal Friend to be either.
It doesn't have to be a Very Strong Top Listed Chess Engine. I prefer that it has my PlayStyle.
A Chess Engine Ranging from 2851-2895 with Good Analyzing Options would Suffice.
Is Sufficient.
( Yes, I am repeating a Variant of that Word. That Word and all of its Variants belongs to my List of Favourite English Words )
but Ofcourse, if you can give the Biggest.. than Give the Biggest ( I can also play against him or her actually )
I found out this List:
http://www.chess.com/downloads/playing+programs+engines
I finally want to get my Games Analyzed ( in WinBoard? Or do you know a better free GUI for this Stuff? ) now that I Book Progress due to Chess Theory ( not Really in the Case of Openings
, I am sorry )What ( or which of this List. It's Big though, you should not Search it for me) Engine that you know, Plays Passively, Defensively, Prophylacticial, Positionally, is not affraid of Draws ( But ofcourse prefers CheckMates huh
), and could be in this List?Ofcourse, when there are Offensive and/or/nor Aggresive ( ( Exchange-)Sacrifices ) chances, one should use those Chances, but I am just not that of a Dynamic, Tactical Player - unlike most others -, Basically, and I don't want my Personal Friend to be either.
It doesn't have to be a Very Strong Top Listed Chess Engine. I prefer that it has my PlayStyle.
A Chess Engine Ranging from 2851-2895 with Good Analyzing Options would Suffice.
Is Sufficient.
( Yes, I am repeating a Variant of that Word. That Word and all of its Variants belongs to my List of Favourite English Words )
but Ofcourse, if you can give the Biggest.. than Give the Biggest
( I can also play against him or her actually )
![[us]](/mwf/v12/flags/us.png "United States")
Buy Rybka and Aquarium.
![[fr]](/mwf/v12/flags/fr.png "France")
For sure, this is the best !!
![[gb]](/mwf/v12/flags/gb.png "United Kingdom")
The problem is this....
All modern engines play prophylactically, but in very concrete ways not general ways.
All modern engines defend superbly but none have a defensive style.
No modern engine plays passively.
No mdoern engine correctly evaluates some passive positions. A classic example is 1.d4 g6, 2.c4 Bg7, 3.e4 d6, 4.Nc3 e5, 5.dxe5 dxe5 6.Qxd8 Kxd8. This is very close to dead level but all engines show quite a big white advantage. Analyse this deeply with an engine or engines and you will see their evaluations gradually drop until eventually they are close to 0 - no engine I have tried can demonstrate a meaningful advantage depsite their high evaluations.
All modern engines get at least the basics of positional play right but some are better than others. Komodo is generally accepted as being very strong positionally - it just puts it's pieces on the right squares and its evaluation is very good. Komodo 4 is commercial but Komodo 3 is still available for fre at the Komodo site. I suggest you try Komodo 3 and if you like it buy Komodo 4 (it's cheap, it's a little stronger than Komodo 3 and IMHO it is good to pay the developers).
But I will make a special mention for Rybka. It is true that Rybka is very tactical. But Rybka 3 was the first engine that I used where I really trusted the evaluation. I feel it taught me quite a bit about chess. If Rybka 3 and I disagree about the eval of a middlegame position and I dig deeper, normally I find Rybka is right - not always but normally. With engines before this when I dug deeper their evaluations were more often wrong - of course they could still be beat me because they were tactically stronger. Its evaluations in endgames are less consistent, but this is problem that all engines have to varying degrees. I don't own Rybka 4. Presumably it is stronger positionally that Rybka 3.
Zappa is also good positionally and has a very balanced playing style across the phases of the game. It is available from the Shredder web site.
Houdini, Critter and Stockfish are good and at least one is stronger than Komodo, Rybka and Zappa at quick time limits. But I do not trust their evaluations and I think Houdini and Critter in particualr rely on superior tactical strength for their higher performance in the rating lists.
With all these engines you will need a GUI to run them in, e.g. Arena (free), Chessbase (expensive) etc.
All modern engines play prophylactically, but in very concrete ways not general ways.
All modern engines defend superbly but none have a defensive style.
No modern engine plays passively.
No mdoern engine correctly evaluates some passive positions. A classic example is 1.d4 g6, 2.c4 Bg7, 3.e4 d6, 4.Nc3 e5, 5.dxe5 dxe5 6.Qxd8 Kxd8. This is very close to dead level but all engines show quite a big white advantage. Analyse this deeply with an engine or engines and you will see their evaluations gradually drop until eventually they are close to 0 - no engine I have tried can demonstrate a meaningful advantage depsite their high evaluations.
All modern engines get at least the basics of positional play right but some are better than others. Komodo is generally accepted as being very strong positionally - it just puts it's pieces on the right squares and its evaluation is very good. Komodo 4 is commercial but Komodo 3 is still available for fre at the Komodo site. I suggest you try Komodo 3 and if you like it buy Komodo 4 (it's cheap, it's a little stronger than Komodo 3 and IMHO it is good to pay the developers).
But I will make a special mention for Rybka. It is true that Rybka is very tactical. But Rybka 3 was the first engine that I used where I really trusted the evaluation. I feel it taught me quite a bit about chess. If Rybka 3 and I disagree about the eval of a middlegame position and I dig deeper, normally I find Rybka is right - not always but normally. With engines before this when I dug deeper their evaluations were more often wrong - of course they could still be beat me because they were tactically stronger. Its evaluations in endgames are less consistent, but this is problem that all engines have to varying degrees. I don't own Rybka 4. Presumably it is stronger positionally that Rybka 3.
Zappa is also good positionally and has a very balanced playing style across the phases of the game. It is available from the Shredder web site.
Houdini, Critter and Stockfish are good and at least one is stronger than Komodo, Rybka and Zappa at quick time limits. But I do not trust their evaluations and I think Houdini and Critter in particualr rely on superior tactical strength for their higher performance in the rating lists.
With all these engines you will need a GUI to run them in, e.g. Arena (free), Chessbase (expensive) etc.
I really like your post, it's one of the best I've seen recently:
True.
True.
Yeah...I think this (all too common) position is a type of fortress. It's not literally a fortress, but it has characteristics of one as all of white's advantages can't be realised easily.
Another common way of getting to this is 1.e4 d6 2.d4 Nf6 (a Pirc...but no) 3.Nc3 e5?! and black has similar ideas.
I wouldn't argue with that and I'm getting very frustrated waiting for 4.1 to come out. However, Richard V hinted that Komodo's strength was more to do with its search than anything else. I think Komodo's evaluation function might be lighter than Rybka's. (but well tuned)
It's also true that Rybka is very aggressive. I have found a unique mathematical way to measure aggression (hopefully this will compliment my Tactical List which I do each month soon) and Rybka is coming up high on the list. HIARCS is also aggressive but I don't always trust its evals. Komodo is low on the list.
I think it is more aggressive, but about the same tactical strength.
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=24848
Not sure about the 'positional' bit but it searches deeper I think - better pruning.
Houdini's evals are weird to begin with but I like what Robert H has tried to do: he's tried to make a score of +1.00 MEAN SOMETHING. In Houdini's case it means 85% chance of victory.
> All modern engines defend superbly but none have a defensive style.
True.
> No modern engine plays passively.
True.
> No mdoern engine correctly evaluates...
> 1.d4 g6, 2.c4 Bg7, 3.e4 d6, 4.Nc3 e5, 5.dxe5 dxe5 6.Qxd8 Kxd8.
Yeah...I think this (all too common) position is a type of fortress. It's not literally a fortress, but it has characteristics of one as all of white's advantages can't be realised easily.
Another common way of getting to this is 1.e4 d6 2.d4 Nf6 (a Pirc...but no) 3.Nc3 e5?! and black has similar ideas.
> Komodo is generally accepted as being very strong positionally
I wouldn't argue with that and I'm getting very frustrated waiting for 4.1 to come out. However, Richard V hinted that Komodo's strength was more to do with its search than anything else. I think Komodo's evaluation function might be lighter than Rybka's. (but well tuned)
> It is true that Rybka is very tactical.
It's also true that Rybka is very aggressive. I have found a unique mathematical way to measure aggression (hopefully this will compliment my Tactical List which I do each month soon) and Rybka is coming up high on the list. HIARCS is also aggressive but I don't always trust its evals. Komodo is low on the list.
> I don't own Rybka 4. Presumably it is stronger positionally that Rybka 3.
I think it is more aggressive, but about the same tactical strength.
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=24848
Not sure about the 'positional' bit but it searches deeper I think - better pruning.
> Houdini, Critter and Stockfish are good and at least one is stronger than Komodo, Rybka and Zappa at quick time limits. But I do not trust their evaluations and I think Houdini and Critter in particualr rely on superior tactical strength for their higher performance in the rating lists.
Houdini's evals are weird to begin with but I like what Robert H has tried to do: he's tried to make a score of +1.00 MEAN SOMETHING. In Houdini's case it means 85% chance of victory.
> No modern engine plays passively.
For some Clearance, I mean Passive in a Relative Sense.
Passive '' good moves '', not passive '' bad moves '' ( but I suppose that you already understood this with Reasonable Sense )
Or is ' Passive ' always in Contrast with ' good '?
Or do you mean Passive Positions?
Anyway, here is a Game where I and my Opponent played ' passive ' moves { the game is still on--the-go by the way )
[Event "Happy Defence"]
[Site "Chess.com"]
[Date "2012.05.06"]
[White "reneminharro"]
[Black "Defence4Gizchehs"]
[Result "*"]
[WhiteElo "1284"]
[BlackElo "1200"]
[TimeControl "1 in 3 days"]
1.d4 d5 2.e3 Nf6 3.Nf3 g6 4.c3 Bg7 5.Qb3 O-O 6.h3 c6 7.g4 Nbd7 8.Rg1 Ne4 9.h4 Qb6 10.Qa3 Re8
11.g5 Qc7 12.Nbd2 Nd6 13.Be2 Nb6 14.Qb3 Be6 15.Qc2 Nf5 16.h5 Rac8 17.hxg6 hxg6 18.Rg2 Nd7 19.e4 dxe4 20.Nxe4 c5
21.dxc5 Nxc5 22.Nfd2 Nh4 23.Rg1 Qh2 24.Rg3 Red8 25.Nf3 Qh1+ 26.Bf1 Nxf3+ 27.Ke2 *
this is Positional Play right?
Do you Consider this as Komodo Style of Play, or better, iis this Komodo Style of Play?
And to keep it fair, here is the Game where I was all Tightned-up already from the Beginning.
[Event "Survive the Guncas assault"]
[Site "Chess.com"]
[Date "2012.05.10"]
[White "flyinguncas"]
[Black "Defence4Gizchehs"]
[Result "*"]
[WhiteElo "1461"]
[BlackElo "1200"]
[TimeControl "1 in 3 days"]
[Variant "Chess 960"]
[SetUp "1"]
[FEN "rbkqbrnn/pppppppp/8/8/8/8/PPPPPPPP/RBKQBRNN w KQkq - 0 1"]
1.d4 d5 2.f4 c6 3.e3 Nf6 4.Ng3 e6 5.Bb4 Rg8 6.c3 Bd7 7.Qe2 Qe8 8.a4 a5 9.Bc5 Bc7 10.Bd3 Kd8
11.Nh5 Ne4 12.Bxe4 dxe4 13.Ng3 f5 14.b4 b6 15.Rb1 axb4 16.Bxb4 g6 17.Ra1 Nf7 18.Qf2 Bd6 19.Qb2 Kc7 20.Bxd6+ Nxd6
What is this Kind of Play?
Certainly not Rybka, it would have gone 5... Bd6 I think.
Also not Houdini, it would have Gone 18... e5 ( provided that it was a Good move. Was this a better move by the way? I didn't like the Tactics and Dynamics that would Arise out of it, but I was a bit Sleepy when I made the move Bd6, and I already like the Defence more. I also want to keep as much Pawns on the Board... But I have a Bad light Squared Bishop! I might be able to let it see the World though, as OnlineChessLessons and/or/nor Jerry from the ChessNetWork would say.. just not give me Advice other than the what-if-History-Pawn-Gambit, the game is still going-on. )
I took my Medicin against the Erysipelas on my Feet at 22:45, London Time ( right? ).
Now I can sleep.
The Possible Download of a new Friend, and a new Combat against her, the Old School Rybka 2.3.2a mp 32-bit, comes Tommorow
![[mx]](/mwf/v12/flags/mx.png "Mexico")
For passive as relative sense, Rybka 2.3.2a is quite passive, I actually think it's the most passive engine of that strength, only surpassed by the passiveness of Naum 3.0. Compare to Rybka 3 Dynamic, now that's an engine that is very offensive when compared to 2.3.2a. And let's not even talk about Thinker, which just plays sacrifices because they're playable, not because they're best.
So on a scale of passiveness to aggressiveness, Rybka 2.3.2a is very close to the passive side, it's very strange that you'd put it on the offensive side.
For chess engines, it's as close as you can get (unless Fritz =>11 is worse), search and you'll see only more aggressive engines of that strength.
Perhaps you're searching for solid engines, in this case Zappa Mexico and Critter =>1.2 are good candidates (but they're not passive).
So on a scale of passiveness to aggressiveness, Rybka 2.3.2a is very close to the passive side, it's very strange that you'd put it on the offensive side.
For chess engines, it's as close as you can get (unless Fritz =>11 is worse), search and you'll see only more aggressive engines of that strength.
Perhaps you're searching for solid engines, in this case Zappa Mexico and Critter =>1.2 are good candidates (but they're not passive).
Well, Than....
'' I will Cease and accept that I was Wrong. ''
I always thought that Rybka 2.3.2a mp 32-bit was Offensive after Reading this:http://www.chessbase.com/newsdetail.asp?newsid=4357 . a Harsch, Tactical, Chess Player Engine.
'... But if it is not, than I will have to Approve the Statement that you made in the Past Tense.
I Quote:
' Also, define "offensive", as, if you make a mistake, all engines will try to punish you. '
'' I will Cease and accept that I was Wrong. ''
I always thought that Rybka 2.3.2a mp 32-bit was Offensive after Reading this:http://www.chessbase.com/newsdetail.asp?newsid=4357 . a Harsch, Tactical, Chess Player Engine.
'... But if it is not, than I will have to Approve the Statement that you made in the Past Tense.
I Quote:
' Also, define "offensive", as, if you make a mistake, all engines will try to punish you. '
That article is rather flawed. I would say that hiarcs is the most aggressive of the 4 programs he names.
WereWolf, now that you are Hopefully still Online, do you know what Parameters are?
I must fill in Parameters for Komodo 3 in my Fresh and still Beloved new Scid GUI ( but I also like ' low Wind on the Board ' and ' Tarrasch's Honor ' ), but I do not exactly know.. what to do, like Igor Smirnov says that.
I did find this however though:
http://www.rybkachess.com/index.php?auswahl=Engine+parameters
I must fill in Parameters for Komodo 3 in my Fresh and still Beloved new Scid GUI ( but I also like ' low Wind on the Board ' and ' Tarrasch's Honor '
), but I do not exactly know.. what to do, like Igor Smirnov says that.I did find this however though:
http://www.rybkachess.com/index.php?auswahl=Engine+parameters
Don't the default values work ok?
Hash size should be at least 512 MB, I'm not sure what you're asking...
Hash size should be at least 512 MB, I'm not sure what you're asking...
I dunno, I want it to have all Perfect you know.
Ah ok, I see now you mean SCID - I don't use that program sorry. Looks interesting though.
> Hash size should be at least 512 MB
Only if the user is going to leave the position being analyzed for several hours. Otherwise, for Interactive Analysis and playing against the engine 128MB are enough.
128MB are enough.
Uly how long does it take your program with your hardware to fill 128megs of hash?
Uly how long does it take your program with your hardware to fill 128megs of hash?
The hash is never "full", when there are hash has contents everywhere, it starts overwriting entries. If the engine is smart at how it's overwritten (this needs to be ticked in some engine's option's like Rybka's "Preserve Analysis") you'll be missing unimportant entries so the benefit of a bigger hash is microscopic. If the engine is dumb at how it's overwritten (like Default Stockfish) then the engine is going to forget the best entries anyway whenever there's a move made, so again, the benefit of bigger hash can't be seen. And if the engine has learning (like Critter 1.4) what is important to keep will be saved to disk, so the amount of hash that engine gets is unimportant (as long as it's high enough, and I think 128MB is high enough.)
It takes 6 seconds for my Critter to show that the hash is 100% full, but there's no penalty for this, and no benefit if it took it more time to "fill it".
It takes 6 seconds for my Critter to show that the hash is 100% full, but there's no penalty for this, and no benefit if it took it more time to "fill it".
![[de]](/mwf/v12/flags/de.png "Germany")
As a rule of thumb doubling hash size adds 5 Elo points playing strength. Vas claimed this years ago - and I checked it and found he was right.
Dadi found otherwise:
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=164647
(First test 64MB topping the list)
(Second test 2048MB was at bottom)
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=164647
(First test 64MB topping the list)
(Second test 2048MB was at bottom)
Dadi's test was very cursory. I suspect that Lukas' test was much more thorough...
I remember when Vas threw out the 5 Elo per hash doubling (in fact, I think I was the one who asked the question). The answer sounded reasonable and I intended to test it, but never got around to it. If anything, I suspect the number is understated for deep depths where the hash has been overwritten by the time the iterative deepening cycle finishes.
I remember when Vas threw out the 5 Elo per hash doubling (in fact, I think I was the one who asked the question). The answer sounded reasonable and I intended to test it, but never got around to it. If anything, I suspect the number is understated for deep depths where the hash has been overwritten by the time the iterative deepening cycle finishes.
> Dadi's test was very cursory. I suspect that Lukas' test was much more thorough...
I believe more in published tests than in "vaporware" tests.
Published tests are frequently not worth the paper they are printed on. We have a site here in the US called PubMed that has millions of medical journal articles. A large percentage are seriously flawed because of poor study design.
The important aspects are:
- Does the study design really test what is being studied to high confidence? and
- Was the study carried out in a methodical manner?
Dadi's test was interesting, and showed the expected high variability, but he didn't run enough samples to come up with valid conclusions (at least based on the published results). I doubt he would argue with this characterization.
Without even seeing Lukas' results, I'm guessing he did a more thorough test, and came out with results consistent with the 5 Elo per doubling hypothesis. Enough to prove the hypothesis to within, say +/- 1 Elo? Almost certainly not, but I'm guessing maybe within +/- 3 Elo. Could I be wrong? Certainly, but from experience, I know that Lukas' results are generally bankable.
The important aspects are:
- Does the study design really test what is being studied to high confidence? and
- Was the study carried out in a methodical manner?
Dadi's test was interesting, and showed the expected high variability, but he didn't run enough samples to come up with valid conclusions (at least based on the published results). I doubt he would argue with this characterization.
Without even seeing Lukas' results, I'm guessing he did a more thorough test, and came out with results consistent with the 5 Elo per doubling hypothesis. Enough to prove the hypothesis to within, say +/- 1 Elo? Almost certainly not, but I'm guessing maybe within +/- 3 Elo. Could I be wrong? Certainly, but from experience, I know that Lukas' results are generally bankable.
And again, I'm going to trust Dadi's test more than your guesses. You're just assuming and you could be wrong.
To me, Lukas's test doesn't exist until I see them. It's a similar concept of "Pics or it didn't happen."
If someone else appeared claiming his tests showed 1 elo improvement per doubling, you would believe him or not believe him depending on who it is? So you believe Lukas's claim just because he is Lukas.
Lukas's tests also showed a much bigger improvement from Rybka 3 to Rybka 4 and the reality was far from it.
What if he tested at hyperbullet time controls? Maybe his results don't hold at longer time controls or analysis, we just don't know as he didn't even disclose the conditions of his test.
To me, Lukas's test doesn't exist until I see them. It's a similar concept of "Pics or it didn't happen."
If someone else appeared claiming his tests showed 1 elo improvement per doubling, you would believe him or not believe him depending on who it is? So you believe Lukas's claim just because he is Lukas.
Lukas's tests also showed a much bigger improvement from Rybka 3 to Rybka 4 and the reality was far from it.
What if he tested at hyperbullet time controls? Maybe his results don't hold at longer time controls or analysis, we just don't know as he didn't even disclose the conditions of his test.
You're just assuming and you could be wrong.
Me? Wrong? Impossible!!!
Reasonable conditions for the test include waiting long enough such that the hash replacement policy is being exercised, but not long enough such that the hash has no useful data remaining after an iterative deepening cycle. This is independent of the tester. Probably a more useful test, but one that would require more instrumentation, would actually measure the hash hit rate as a function of hash size, for a number of positions along with time to solution.
What if he tested at hyperbullet time controls?
This would be a silly test if the hash replacement policy was not exercised.
So you believe Lukas's claim just because he is Lukas.
Yes! Life is short, so it's important to have trusted sources. Nobody has time to verify everything they read or hear. Most people, for example, trust their doctors. I don't trust my doctors and require corroboration in the literature, but for something less important like chess engine speed, I am happy to trust respected sources.
Me? Wrong? Impossible!!!
Reasonable conditions for the test include waiting long enough such that the hash replacement policy is being exercised, but not long enough such that the hash has no useful data remaining after an iterative deepening cycle. This is independent of the tester. Probably a more useful test, but one that would require more instrumentation, would actually measure the hash hit rate as a function of hash size, for a number of positions along with time to solution.
What if he tested at hyperbullet time controls?
This would be a silly test if the hash replacement policy was not exercised.
So you believe Lukas's claim just because he is Lukas.
Yes! Life is short, so it's important to have trusted sources. Nobody has time to verify everything they read or hear. Most people, for example, trust their doctors. I don't trust my doctors and require corroboration in the literature, but for something less important like chess engine speed, I am happy to trust respected sources.
Well, if you believed Lukas's expected improved elo from Rybka 3 to Rybka 4 I have a bridge to sell you.
Please provide a reference. Was this based on testing, or was he passing on news from Vas? And are you sure it wasn't the number he quoted before Vas dumbed it down? And what hardware was it for?
Wait, you're a beta tester, you can look yourself:
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=17704 (only beta testers can see this link)
That performance amounts to 3230 elo in the CCRL, far away from truth.
Please note I don't doubt Lukas results, he really went and played the games and got those results. What I'm saying is those results weren't meaningful in saying what one expects to get in real life, so I doubt also the 5 elo doubling would apply in real life, unless he discloses conditions, and they're meaningful.
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=17704 (only beta testers can see this link)
That performance amounts to 3230 elo in the CCRL, far away from truth.
Please note I don't doubt Lukas results, he really went and played the games and got those results. What I'm saying is those results weren't meaningful in saying what one expects to get in real life, so I doubt also the 5 elo doubling would apply in real life, unless he discloses conditions, and they're meaningful.
he really went and played the games and got those results.
Yes. He ran a test and got those results. That's all you can ask. Maybe the result of the test was an outlier. This happens a small percentage of the time. Or maybe there is was a very large difference with Stockfish 1.7.1 on Lukas' hardware.
That performance amounts to 3230 elo in the CCRL, far away from truth.
I call bullshit here. Beating up on one engine, Stockfish 1.7.1 in this case, doesn't get you a 3230 Elo rating in CCRL.
What I'm saying is those results weren't meaningful in saying what one expects to get in real life
Those results are meaningful in saying what you would expect to get with those exact engines on that exact hardware with the exact book. It's not Lukas' fault if you extrapolate the results incorrectly.
so I doubt also the 5 elo doubling would apply in real life, unless he discloses conditions, and they're meaningful.
It's not important whether you accept his results, or not. It's important whether the result is meaningful or not. I suspect that Lukas has done more testing of this than you or I, and he has no reason to be biased, so I accept his results. Others can make up their own minds about whether they want to accept this result from someone who ran the test, or from someone who didn't.
Yes. He ran a test and got those results. That's all you can ask. Maybe the result of the test was an outlier. This happens a small percentage of the time. Or maybe there is was a very large difference with Stockfish 1.7.1 on Lukas' hardware.
That performance amounts to 3230 elo in the CCRL, far away from truth.
I call bullshit here. Beating up on one engine, Stockfish 1.7.1 in this case, doesn't get you a 3230 Elo rating in CCRL.
What I'm saying is those results weren't meaningful in saying what one expects to get in real life
Those results are meaningful in saying what you would expect to get with those exact engines on that exact hardware with the exact book. It's not Lukas' fault if you extrapolate the results incorrectly.
so I doubt also the 5 elo doubling would apply in real life, unless he discloses conditions, and they're meaningful.
It's not important whether you accept his results, or not. It's important whether the result is meaningful or not. I suspect that Lukas has done more testing of this than you or I, and he has no reason to be biased, so I accept his results. Others can make up their own minds about whether they want to accept this result from someone who ran the test, or from someone who didn't.
> It's not Lukas' fault if you extrapolate the results incorrectly.
But since you don't know what were the conditions of his tests, you can't know if he's extrapolating incorrectly (maybe he tested for game in one second, game in 10 seconds, game in 20 seconds, game in 30 seconds, and game in 1 minute and got 5 elo per doubling and extrapolated, we just don't know).
> I suspect that Lukas has done more testing of this than you or I, and he has no reason to be biased, so I accept his results.
But his results may only be true for whatever unknown conditions he used, and not for the general case. So it's quite possible for Interactive Analysis and playing chess against the engine, that 128MB are enough and that doubling wouldn't give 5 elo more.
Also, all the tests I've seen Lukas doing were on bullet or fast blitz time controls, never on standard time controls, so I see no reason to assume that the doubling hash testing was done in the right time controls.
>That performance amounts to 3230 elo in the CCRL, far away from truth.
Conditions were totally different from CCRL conditions (bullet games, 6 cores, 8 book moves...) - so you can't compare results. And the first screenshot was after only 170 games.
>so I doubt also the 5 elo doubling would apply in real life, unless he discloses conditions, and they're meaningful
The setup was simple - I took Rybka with lowest possible hash (2MB) and played lots of games vs. Rybka with much more hash (512 MB). Games were 3 min + 0, the result was around + 40 Elo.
Btw. hash usage heavily depends on the hardware you have. If I set up Critter with 4 GB of hash, it takes less than 20 sec. to get a hashfull indication of > 50% on a 2x Xeon E5-2680. On a smaller computer it takes much longer.
Thanks for revealing conditions.
Now let's see if Banned for Life considers them meaningful. You didn't even test 4MB, 8MB, 16MB, 32MB, 64MB or even if there's a difference between 128MB and 512MB.
You only tested 2MB Vs. 512MB at 3 + 0, got +40 ELO and extrapolated. You don't even know if the result would be the same with 4MB Vs. 1024MB. I find Dadi's test more meaningful.
Now let's see if Banned for Life considers them meaningful. You didn't even test 4MB, 8MB, 16MB, 32MB, 64MB or even if there's a difference between 128MB and 512MB.
You only tested 2MB Vs. 512MB at 3 + 0, got +40 ELO and extrapolated. You don't even know if the result would be the same with 4MB Vs. 1024MB. I find Dadi's test more meaningful.
I consider the test Lukas did to be very meaningful. Could it have been more extensive? Sure, but Lukas' test isn't based on a very limited sample size, and shows that under some circumstances a larger hash outperforms a smaller hash. After coming to this conclusion, there is a need to quantify the results with different test conditions.
As far as a hypothesis for the Elo difference between 4MB and 1024MB, I wouldn't be surprised if it were slightly below 40 (i.e. there might be more advantage in going from 2MB to 4MB than from going from 512MB to 1024MB for 3 minute games on fast hardware).
Dadi's test showed that for a particular position, increasing hash size doesn't always improve performance (time to solution). This is interesting, but in no way indicative of statistical expectations from running large numbers of positions with different hash sizes. If you can't tell the difference between these two situations, you would make a very poor scientist (or maybe you should be a doctor, because there are a high percentage of studies where doctors form conclusions from ridiculously small sample sizes).
As far as a hypothesis for the Elo difference between 4MB and 1024MB, I wouldn't be surprised if it were slightly below 40 (i.e. there might be more advantage in going from 2MB to 4MB than from going from 512MB to 1024MB for 3 minute games on fast hardware).
Dadi's test showed that for a particular position, increasing hash size doesn't always improve performance (time to solution). This is interesting, but in no way indicative of statistical expectations from running large numbers of positions with different hash sizes. If you can't tell the difference between these two situations, you would make a very poor scientist (or maybe you should be a doctor, because there are a high percentage of studies where doctors form conclusions from ridiculously small sample sizes).
> Dadi's test showed that for a particular position
What particular position? He played 1600 games with a book, and a Round Robin where all the hash sizes have to play one another. He also used increment that will make games less random as the Lukas's games may have been decided because one engine didn't have enough time to play the best moves.
All Lukas's tests show is that 512MB get +40 elo to 2MB in 3 0 time control, nothing else can be concluded (it's possible improvement is not linear). He didn't check if the improvement from 2MB to 8MB was 10 elo and if the improvement from 128MB to 512MB was 10 elo. His tests were inconclusive.
I guess you are referring to this:
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=164647#pid164647
In the first test, 1'1" games on 1 core probably don't require anything more than a small hash, so the results are not surprising. In the second test, there is a penalty associated with large hashes when large pages are not being used, so once again, it's not surprising that the largest hashes perform poorly in that test. It's also the case that there is only so much hash required for a 1 core 10'1" game. But in the second test, there is a clear correlation between larger hash size and better performance, up to a point.
So Dadi's results are not at all surprising. If this isn't apparent to you, it's because you lack understanding of the theoretical advantages and disadvantages of a various hash sizes, related to the product of time per move and engine speed.
Lukas' test was more relevant since it used a higher time-processing product and since it used hash values that were reasonable for that product. Unfortunately, you are too obstinate to grasp the importance of the test, but it wasn't that doubling the hash always produced a gain, or that the gain was 5 Elo per doubling, but that increasing hash size under certain circumstances produces a measurable increase in strength.
http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=164647#pid164647
In the first test, 1'1" games on 1 core probably don't require anything more than a small hash, so the results are not surprising. In the second test, there is a penalty associated with large hashes when large pages are not being used, so once again, it's not surprising that the largest hashes perform poorly in that test. It's also the case that there is only so much hash required for a 1 core 10'1" game. But in the second test, there is a clear correlation between larger hash size and better performance, up to a point.
So Dadi's results are not at all surprising. If this isn't apparent to you, it's because you lack understanding of the theoretical advantages and disadvantages of a various hash sizes, related to the product of time per move and engine speed.
Lukas' test was more relevant since it used a higher time-processing product and since it used hash values that were reasonable for that product. Unfortunately, you are too obstinate to grasp the importance of the test, but it wasn't that doubling the hash always produced a gain, or that the gain was 5 Elo per doubling, but that increasing hash size under certain circumstances produces a measurable increase in strength.
> but it wasn't that doubling the hash always produced a gain, or that the gain was 5 Elo per doubling
That's what Lukas claimed (that the gain would be 5 elo per doubling). It seems we're on agreement that Lukas's test doesn't show that doubling the hash produces 5 elo per doubling. That's what I was claiming, that Lukas's tests weren't meaningful for that, if they are meaningful to show something else, then I don't dispute that.
Vas claimed a gain of 5 Elo per doubling of hash size. This number didn't come from his tests, but was a theoretical value resulting from the way he implemented hash in Rybka.
I wanted to find out how realistic his assumptions were. It requires many thousands of games to find tiny Elo differences. So it's not very clever testing all hash sizes. So I chose a big jump in hash size just to get some easy to find difference in playing strength. My results perfectly matched Vas' theory. Nothing more - and nothing less.
I assume there is a ceiling effect for the usefulness of hash size which depends on CPU speed, number of cores and speed of game play. But my overall conclusion is: it's a good idea to use lots of hash.
I wanted to find out how realistic his assumptions were. It requires many thousands of games to find tiny Elo differences. So it's not very clever testing all hash sizes. So I chose a big jump in hash size just to get some easy to find difference in playing strength. My results perfectly matched Vas' theory. Nothing more - and nothing less.
I assume there is a ceiling effect for the usefulness of hash size which depends on CPU speed, number of cores and speed of game play. But my overall conclusion is: it's a good idea to use lots of hash.
> But my overall conclusion is: it's a good idea to use lots of hash.
If you're fighting against an opponent with low hash on short time controls. Nothing else has been tested. If you're fighting against an opponent at 1 + 1 sec time control giving GBs of hash to the engine may actually hurt.
Also, please note the original question was for the Komodo chess engine, not for Rybka.
> Vas claimed a gain of 5 Elo per doubling of hash size. This number didn't come from his tests, but was a theoretical value resulting from the way he implemented hash in Rybka.
How this can be a theoretical value, beats me.
It certainly seems to be a very first assertion of the kind...
(as much as i respect Vas...)
Never came across such thing in the computer chess literature...
How do you find old topics?
It is always a question for me.
It is always a question for me.
![[fi]](/mwf/v12/flags/fi.png "Finland")
> That article is rather flawed. I would say that hiarcs is the most aggressive of the 4 programs he names.
I thought Junior was more aggressive than Hiarcs?
> I thought Junior was more aggressive than Hiarcs?
Maybe the latest version is. But back then Junior excelled in positions where there was material imbalance, it wasn't always that aggressive IMO. In fact I think it was you who (rightly) claimed HIARCS Padderborn to be a really aggressive engine....my tests confirm this.
> Maybe the latest version is.
People have claimed that Junior has been losing its aggressiveness in recent versions, saying Junior 7 was the peak of aggressiveness.
Please note that I'm not making any claims, I'm asking questions, if you have measured Junior's aggressiveness and it came out low, it means people regarding Junior as aggressive in the past were giving subjective opinions that were wrong. I find it very interesting because it seems people were perceiving something that wasn't there.
You may be right about Junior - I don't have the earlier versions. I have J10 & J12.5. I've run a series of tests to measure aggressiveness. (a bit like a tactical test suite except instead of measuring tactics I'm trying to measure tendency to attack the enemy king)
Junior 12.5 is coming out 'moderately aggressive' and 'strongly preferring material imbalance'.
Hiarcs PB 2007 is very aggressive, subsequent versions less so. (although elo rises with later versions which masks this)
Rybka 4.1 is fairly high on the list, Houdini a little below.
Komodo is low on the list and version 4 is less likely to attack the king than v3
Junior 12.5 is coming out 'moderately aggressive' and 'strongly preferring material imbalance'.
Hiarcs PB 2007 is very aggressive, subsequent versions less so. (although elo rises with later versions which masks this)
Rybka 4.1 is fairly high on the list, Houdini a little below.
Komodo is low on the list and version 4 is less likely to attack the king than v3
Thanks Werewolf, I think it's great that someone is making objective tests on engines' playing style. No longer can someone say "Junior is very aggressive" out of thin air.
Well, why Buying if a Rating of 2851-2895 would Suffice?
And Rybka is a Offensive, Tactical Engine Right?
or maybe not... anyway, Rybka 2.3.2a mp 32-bit is offensive, and I don't want to Buy Rybka just like that anyway. Also because I don't have the right Personal Situation for it.
please give a better answer.
And Rybka is a Offensive, Tactical Engine Right?
or maybe not... anyway, Rybka 2.3.2a mp 32-bit is offensive, and I don't want to Buy Rybka just like that anyway. Also because I don't have the right Personal Situation for it.
please give a better answer.
> Well, why Buying if a Rating of 2851-2895 would Suffice?
Indeed, I don't think you need a commercial engine. I suggest Critter 1.4:
http://www.vlasak.biz/critter/
>And Rybka is a Offensive, Tactical Engine Right?
All engines are tactical, though it can be argued Komodo is less tactical. Also, define "offensive", as, if you make a mistake, all engines will try to punish you.
>or maybe not... anyway, Rybka 2.3.2a mp 32-bit is offensive, and I don't want to Buy Rybka just like that anyway.
All engines with rating >2850 are more offensive than Rybka 2.3.2a.
>All engines with rating >2850 are more offensive than Rybka 2.3.2a.
so we agree that Hiarcs 14 will be more offensive than Rybka 2.3.2a?
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill