Not logged inRybka Chess Community Forum
1 2 3 4 5 6 7 8 Previous Next  
Parent - By lkaufman (*****) Date 2012-03-09 16:22
That may have been a genuine innovation, I don't know.
Parent - - By Richard Vida (**) Date 2012-01-06 00:48 Edited 2012-01-06 01:01
You are playing a "saint" here. Sure you wouldn't want someone digging into Komodo and compare side-by-side its search with Ippo?

[to the public: contrary to the hype, Komodo's eval is not too heavy (in fact much lighter than Rybka3). It is strong mainly because of a very good search algorithm. And it is slow not because of knowledge, it is slow because of sloppy coding - especially the 32bit version which is almost 100 elo weaker than 64bit version]
Parent - By sarciness (***) [gb] Date 2012-01-06 04:39
Ouch, that is hard criticism.
Parent - By Ray (****) Date 2012-01-06 05:47 Edited 2012-01-06 05:57

> You are playing a "saint" here. Sure you wouldn't want someone digging into Komodo and compare side-by-side its search with Ippo?


Agreed with the "saint" bit. Even if Komodo program is "original", is it ethically acceptable that Don has studied Ippo and its predecessors, code that should never have been open to public eyes ? I'm getting quite sick of this "we are whiter than white" attitude from the Komodo team. They are not. But in the new world of computer chess, all is OK. What I am saying is: do study Ippo. You'd be stupid not to. But then don't dare claim the moral high ground.
Parent - By Werewolf (*****) [gb] Date 2012-01-06 08:46

> [to the public: contrary to the hype, Komodo's eval is not too heavy (in fact much lighter than Rybka3). It is strong mainly because of a very good search algorithm. And it is slow not because of knowledge, it is slow because of sloppy coding


*CRASH* and everyone does a double-take and wonders if they misread your post. Could you explain how you know Komodo has a lighter eval than Rybka 3? I have noticed that Komodo is weak in the area of King Safety, it doesn't seem to know when there are chances against the king in the same way HIARCS etc do.

Its tactical performance isn't great either, but you say its search is good, so I assume you mean its non-tactical search.
Parent - - By lkaufman (*****) Date 2012-01-06 16:37
We would welcome a comparison of Komodo and Ippo and other top programs. I'm sure that anyone would quickly see that our search is more like Stockfish than like Ippo.
     You are quite correct about the 32bit Komodo. It is whatever the compiler spit out. Don did not write any code specific for the 32 bit version. Not to say he won't do so in the future though.
Parent - - By tomgdrums (****) Date 2012-01-06 18:03

> We would welcome a comparison of Komodo and Ippo and other top programs. I'm sure that anyone would quickly see that our search is more like Stockfish than like Ippo.
>      You are quite correct about the 32bit Komodo. It is whatever the compiler spit out. Don did not write any code specific for the 32 bit version. Not to say he won't do so in the future though.


With all due respect Larry, I am finding Komodo to be VERY MUCH (identical!) like Hoduini in its move choices when I do post game analysis.  Komodo 4 has yet to show an alternative move, while so-called weaker engines such as Hiarcs 13.2 and the ever exciting Smarthink 1.2 QUITE OFTEN suggest very interesting alternatives to Houdini!  And these alternatives are always corroborated by Houdini as being very valid if ever so slightly "weaker" (which is okay for OTB play!)

Why does Komodo, which is supposed to be different, not give different analysis and move choices.  Critter 1.4 also gives different choices than Houdini.

It seems that Hiarcs is much better for OTB post game analysis than Komodo 4.

I really do think you guys got way too obsessed with chasing Houdini instead of coming up with a really good and interesting engine.  I will keep using Komodo 4 and maybe I will be surprised but so far I am seeing Komodo as less of an alternative option for analysis.
Parent - - By lkaufman (*****) Date 2012-01-06 18:30
In my own analysis for my book I did not notice any particular correlation between Houdini and Komodo's move choices. Perhaps you are talking about tactical positions where there really is a "best" move. Any similarity in non-tactical moves can only be due to the fact that I was responsible for the eval tuning of Rybka 3 as well as Komodo, and Ippo's eval is very close to Rybka 3, and Houdini was at least originally identical to Ippo. Komodo and Rybka 3 were tuned completely independently, but there is going to be some similarity as the optimum values probably don't differ drastically based on the differences in the search. Also we only used the more clearly beneficial terms (but with different definitions in the non-trivial cases) in Komodo, as did Ippo, so Komodo may actually have a closer eval to Ippo than to Rybka for that reason. So if you like, you can say that both Komodo and Houdini use evals based on my work; that would be the similarity.
Parent - - By Stonehenge (***) Date 2012-01-06 19:27

> I was responsible for the eval tuning of Rybka 3 as well as Komodo ...
> ... Komodo and Rybka 3 were tuned completely independently, but there is going to be some similarity as the optimum values probably don't differ drastically based on the differences in the search.
> ... Komodo may actually have a closer eval to Ippo than to Rybka for that reason


1) How ethical is it to change job and take the company secrets to your new employer to create a commercial competitor?
Shouldn't the Rybka team sue you for reusing Rybka know-how in Komodo?

2) A good summary of what you're saying is that Komodo combines Stockfish search with IPPOLIT or Rybka evaluation.
100% original obviously... :eek:
Parent - - By lkaufman (*****) Date 2012-01-06 20:09
Actually Vas said he approved of my working on Komodo, so that my skills would not atrophy in case I later went back to work on Rybka. It's not a question of "company secrets"; I had certain ideas about eval that were used in Rybka, and I never promised (in writing or verballly) not to use my own ideas in another program someday. In any case it all became moot when Robolito came out with the entire eval made public.
     As for the search, we started with one similar to Stockfish, which uses a fairly standard type of search with some very good pruning ideas such as looking at less and less moves towards the end nodes. I don't know whether SF came up with this on their own or got the idea from some other program. I knew that both Rybka and SF did this even before Robolito came out, and others knew about it because of Strelka. But we have added many new and original ideas to the search, which is why Komodo is now clearly stronger than Stockfish, Robolito, and Rybka, though eval ideas also help.
     Finally, if we want to talk about who really deserves the most credit for the recent explosion in strength of chess computers, I propose the name of M.I.T. Professor McCallister.
Parent - - By Stonehenge (***) Date 2012-01-06 20:59

> As for the search, we started with one similar to Stockfish, ...


Thank you for pointing out that Komodo started as "similar to Stockfish".
You sure made Richard's point: On the other hand Komodo(Doch) came out as a 3000 engine out of the blue... Now who used something as a "template"?
Parent - - By lkaufman (*****) Date 2012-01-06 21:19
But the Stockfish search was fairly standard, so it's like saying that the search was similar to Fruit or Crafty except with some recent ideas added. I don't mean to imply that Stockfish had no original ideas, but they were mostly details, I'm talking about the big picture. I think the SF guys will agree that Stockfish is a fairly conventional program in general (i.e. no huge revolutionary idea, just lots of very good details), whereas Rybka and the Ippos are rather different and immediately recognizable in many ways.
Parent - By Stonehenge (***) Date 2012-01-06 21:30

> But the Stockfish search was fairly standard, so it's like saying that the search was similar to Fruit or Crafty except with some recent ideas added. I don't mean to imply that Stockfish had no original ideas, but they were mostly details, I'm talking about the big picture. I think the SF guys will agree that Stockfish is a fairly conventional program in general (i.e. no huge revolutionary idea, just lots of very good details), whereas Rybka and the Ippos are rather different and immediately recognizable in many ways.


ROFL. Stockfish a fairly conventional program in general, without original ideas other than some details?
And you just took the "standard, nothing revolutionary" stuff from Stockfish to start with Komodo?
Parent - - By tomgdrums (****) Date 2012-01-06 21:14

> In my own analysis for my book I did not notice any particular correlation between Houdini and Komodo's move choices. Perhaps you are talking about tactical positions where there really is a "best" move. Any similarity in non-tactical moves can only be due to the fact that I was responsible for the eval tuning of Rybka 3 as well as Komodo, and Ippo's eval is very close to Rybka 3, and Houdini was at least originally identical to Ippo. Komodo and Rybka 3 were tuned completely independently, but there is going to be some similarity as the optimum values probably don't differ drastically based on the differences in the search. Also we only used the more clearly beneficial terms (but with different definitions in the non-trivial cases) in Komodo, as did Ippo, so Komodo may actually have a closer eval to Ippo than to Rybka for that reason. So if you like, you can say that both Komodo and Houdini use evals based on my work; that would be the similarity.


Your tactical position vs positional position argument doesn't make sense.  Hiarcs and Smarthink and Critter ALL have found viable alternatives to Houdini moves on their own while Komodo has not.

The other engines when given time will usually agree that Houdini's move is "best" but the margin is often so slim that all the options are viable OTB thus giving the user more ideas to try when playing actual games.  This is what Uly discusses below.

I purchased Komodo for a more "positional" opinion when doing post game analysis.  So far it has let me down as it gives no different opinion.  Houdini, Critter, Hiarcs and Smarthink are viable and differ often enough to give me food for thought.  So far,, Komodo is just like analyzing with Houdini.
Parent - - By oudheusa (*****) [ch] Date 2012-01-06 21:19
I do not share your experience; I find Komodo to to significantly make different move choices than Houdini or Rybka.
Parent - By tomgdrums (****) Date 2012-01-06 22:23

> I do not share your experience; I find Komodo to to significantly make different move choices than Houdini or Rybka


I wish my experience was false for me.  I really wanted Komodo to be an interesting engine for post game analysis but am so far finding it not very useful at all because it does not provide any original moves.  (only once so far)  Hiarcs is proving to (once again) to be a more useful engine as a compliment to Houdini and Critter.
Parent - - By NATIONAL12 (Gold) [gb] Date 2012-01-06 23:13
i beg to disagree with you on K4 moves being very similar to H2.
see my testing at 40/40.http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=23839;pg=1
Parent - - By tomgdrums (****) Date 2012-01-07 00:20

> i beg to disagree with you on K4 moves being very similar to H2.
> see my testing at 40/40.http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?tid=23839;pg=1


That is just an ongoing run down of your tournament.  (which I had already been following...in fact it was one of the reasons I purchased Komodo).

But there isn't a whole of discussion of move choices and in my standard games this week where I have done post game analysis, Komodo is not picking alternate moves from Houdini.  Only once has it done so.  Hiarcs and Critter are picking alternate moves more than Komodo.  I am not arguing Komodo's strength, I am worried that it is not a useful analysis partner as it does not seem to be giving much alternate information. 

I haven't completely giving up on it yet but given Larry's boasting of how great an analysis partner Komodo 4 would be AND his rather passive-aggressive bashing of Critter as "related" to Houdini while Komodo was OH so different, I expected a LOT MORE from Komodo for post game analysis.   But I should have remembered that the BUYER should BEWARE!

Sorry but that is the truth as it stands now.
Parent - - By NATIONAL12 (Gold) [gb] Date 2012-03-07 22:15
i admit this link only gives games but i watched about a third of them and it suprised me that ponder hits were much lower than i expected.
Parent - By tomgdrums (****) Date 2012-03-08 03:31

> i admit this link only gives games but i watched about a third of them and it suprised me that ponder hits were much lower than i expected.


Actually after much more use, I like Komodo 4 as an analysis partner with Houdini.  It is different enough to give me some choices.
Parent - By Arrière Pensée (Gold) Date 2012-01-07 00:23
Perhaps you should post some games. Just go into the Chessbase folder and pull out the games in doc's.
Parent - By Razor (****) [gb] Date 2012-03-09 05:38
Agreed - Hiarcs 13.2 is a valuable part of my toolkit and used as much as any other engine in my toolkit - ranking lists alone are not enough, how an engine plays is far more important and on that score Mark has done an excellent job over the many years in this game - looking forward to the next version.
Parent - - By Uly (Gold) [mx] Date 2012-01-06 20:26

> I'm sure that anyone would quickly see that our search is more like Stockfish than like Ippo.


The user doesn't care about that. What the user cares about is the end result, what move does the engine play in the end. If Komodo picks moves of Houdini more often than Critter picks moves from Houdini in real life positions (it doesn't matter if they're tactical or positional, but that they happen in the positions users analyze) then related to Houdini Critter seems more original than Komodo.

I'm sad of seeing these "my engine is more original than yours" discussions, specially when the top engines disagree more with each other yet suggest workable plans more than ever (at least if you remove Houdini from there).
Parent - - By lkaufman (*****) Date 2012-01-06 20:37
Could you explain your remark about removing Houdini?
Parent - - By Uly (Gold) [mx] Date 2012-01-06 21:45
The overlap in move choice with Houdini and other engines seems bigger than the overlap in move choice between other engines among one another. I'll try to explain with a graphic.


Note: It's just a visual concept with sizes greatly exaggerated not to any conceivable scale. Not shown is the overlap between Yellow and Blue, nor overlaps in where 3 circles meet.

So here, suppose Red is the original moves of Houdini, Yellow the original moves of Critter, Cyan the original moves of Rybka, and in Blue the ones from Komodo. And it can be seen that there's a greater overlap between Houdini and the other engines than with engines to each other, so that if you removed Houdini, the overlap would go down. That's exactly what I do in my analysis because I know the plan Houdini would suggest will be probably suggested by another engine, the big overlaps of Houdini are not worth the time IMO.

And the original moves of Houdini usually are worse that what I already built with other engines. I still find it odd that Houdini is so good at beating other engines in the rating lists and its superb elo while I find Rybka and Critter (and even Naum 4.2...) significantly better for analysis, though things change when you analyze with many engines, as the positions where the top ones have flaws are more obvious.
Parent - - By Uly (Gold) [mx] Date 2012-01-06 22:15
Wait, I got it, maybe such overlap is the reason Houdini is so good for games, suppose the best move is in the orange area, then Rybka and Komodo would miss it. If the best move is in the grey area, Critter and Komodo will miss it. If it's in the purple area, Rybka and Komodo will miss it. But in these instances Houdini would catch them all. The instances in where the best move is in the green area or in the Cyan-Blue are much rarer.

If moves in where only one engine finds the best move are rare, and overlaps are more common, Houdini would be better in finding the best moves because those areas are bigger on the graphics, and it doesn't matter if other engine would find that move as well, for games (but for analysis, overlap is bad).

Or maybe it's just balderdash, I just found the thought interesting.
Parent - By tomgdrums (****) Date 2012-01-06 22:21

> Wait, I got it, maybe such overlap is the reason Houdini is so good for games, suppose the best move is in the orange area, then Rybka and Komodo would miss it. If the best move is in the grey area, Critter and Komodo will miss it. If it's in the purple area, Rybka and Komodo will miss it. But in these instances Houdini would catch them all. The instances in where the best move is in the green area or in the Cyan-Blue are much rarer.
>
> If moves in where only one engine finds the best move are rare, and overlaps are more common, Houdini would be better in finding the best moves because those areas are bigger on the graphics, and it doesn't matter if other engine would find that move as well, for games (but for analysis, overlap is bad).
>
> Or maybe it's just balderdash, I just found the thought interesting.


I don't think that is balderdash!   Maybe that is why I like using Houdini for post game analysis.  It is a great universal engine and I generally understand the move choices.  I do like having a few engines to check interesting or crucial positions as it can help with post game analysis.
Parent - - By tomgdrums (****) Date 2012-01-06 22:18
I keep forgetting to include Naum 4.2 in my post game analysis!! :confused:
Parent - By Uly (Gold) [mx] Date 2012-01-07 16:29
I like Zappa better than Naum, it's just that I seem to be the only guy around that still uses Naum, I think that it's probably the best engine with the style of engines from the pre-Rybka 3 era.
Parent - By Werewolf (*****) [gb] Date 2012-01-09 07:18

> [to the public: contrary to the hype, Komodo's eval is not too heavy (in fact much lighter than Rybka3). It is strong mainly because of a very good search algorithm. And it is slow not because of knowledge, it is slow because of sloppy coding - especially the 32bit version which is almost 100 elo weaker than 64bit version]


Please can you explain this.
Parent - By Razor (****) [gb] Date 2012-03-08 05:55
Your statement (1) is written as a fact; I was under the impression that statements of facts have supporting evidence - do you have this evidence and if so, can this be published here?
Parent - - By Arrière Pensée (Gold) Date 2012-01-05 20:45
My suggestion is to stop doing this! It doesn't look good. :wink: Richard has a very strong following and after a while it won't sit well with them.
Parent - By cipri (**) [de] Date 2012-01-05 20:52
Yes, it seems you care about "looking good", and not about the true.

Why are you not again explaining why Houdini 2.0 doesnt contain any code of Ippolit, like you did before?
You forgot you great argument: "if houdini 2.0 would contain any code from ippolit/fruit... then there would be not promotion for Houdini on this site".
Did you remember how you "verbally attacked" dr. hyatt because he said that houdini uses ippolit code?
I guess if you have any kind of dignity, you should do the same with larry what you did to dr. hyatt.
Parent - By Arrière Pensée (Gold) Date 2012-01-05 22:02
:yell:

Ignore is a beautifully thing!
Parent - - By Richard Vida (**) Date 2012-01-06 00:39

> The similarity and the use of nearly all Ippo ideas is not a secret


Neither is secret that almost all those ideas are also used by Komodo. Except a few, which you either don't understand well or just can not make it work.

> while Vida wrote his own program using Ippo as a "template"


False. I used no template. I wrote my engine entirely on my own, long before even the name Komodo existed. First version of Critter that anybody cared enough about to test was v0.38 (about 2500 CCRL scale). I went through all the pains implementing stuff like LMR/razoring/null move etc, etc gradually improving with each release. Tried different approaches and settled with what seemed to work best. On the other hand Komodo(Doch) came out as a 3000 engine out of the blue... Now who used something as a "template"?

I understand that you have commercial interests... But I am sure you will be much more popular within the community if you actually competed with a superior product rather than just hype and badmouthing other competitors. You just accused me of "cloning" although you said it with much softer wording. I sent you my sources in good faith wanting to stir up the competition at the top, never wanted anything in return. Please try to _compete_ instead of belittling other's work...

> he has done nothing wrong and I wish him success in besting Houdart


I wish success to everyone, including you.
Parent - By Banned for Life (Gold) Date 2012-01-06 01:25
A very impressive argument. Clean and sharp. Salute!

But unlike some of us :twisted:, Larry doesn't have a mean bone in his body. He's just repeating what he hears from the community. Now that you've set him straight, I'm sure he'll seek the truth, which shouldn't be too difficult with the source that you've provided.

Looking forward to great things from Critter!
Parent - - By sarciness (***) [gb] Date 2012-01-06 04:46
A classy reply.

I have not yet tried Critter so I can't judge its merits yet, but I think I'll go download it now!
Parent - By Uly (Gold) [mx] Date 2012-01-06 14:57
I find it better for analysis than Rybka 4.1 or Houdini (though the latter goes without saying.)
Parent - By Nelson Hernandez (Gold) [us] Date 2012-01-06 11:13
Refreshing!  Admirable candor.
Parent - - By lkaufman (*****) Date 2012-01-06 16:33
1. Komodo does not use any Ippo formulas, parameters or the like, with the possible exception of one high-depth pruning formula which was inspired by Ippo but which differs considerably in the particulars. We don't even use the basic Ippo (really Rybka) idea of "expected gain", as we tried it but didn't find it helpful. We do of course use general ideas, ideas that can be expressed in simple English without mentioning exact numbers. For example "Singular Extension of Hash Moves". Our lower-depth pruning is similar to Stockfish, not to Ippo. So it is incorrect to say that Komodo uses "almost all" Ippo ideas.
     2. My template comment was based partly on your own statement that your Pascal program was "heavily Ippo-ized" or something like that. Perhaps I was wrong to assume that this also applied to how your main program came into being. Did you write your main program before or after the release of Robbolito? If you tuned LMR etc. yourself, why are the values you use so similar to or identical to Ippo in so many cases? We have never found Ippo values to be good ones. As for Doch, remember we had top level chess programs on the market twenty years ago. Also, the early Doch predated the Ippo release.
     3. Since you were kind enough to send us Critter sources a while ago, I'd like to tell you something in return that should help you improve Critter. I ran some experiments together with Mark Leffler in which we tried turning off various ideas and formulas in one of the Ippo versions, and we were surprised to discover that in many cases the ideas or formulas were worthless or worse. I think you can improve Critter by taking those formulas and ideas that are identical (or nearly so) to Ippo (and unique to Ippo) and greatly modifying them or removing them. My opinion now is that Ippo is not a very good program from the algorithmic standpoint, it is just very fast perhaps due to excellent coding. The proof is that it is little if any stronger than Stockfish, which in general uses common-sense algorithms for pruning, not obscure ideas. This is probably why we have been able to pass Ippo despite a large speed handicap.
Parent - - By Richard Vida (**) Date 2012-01-08 07:01 Edited 2012-01-08 07:12

> We don't even use the basic Ippo (really Rybka) idea of "expected gain", as we tried it but didn't find it helpful


SF uses it too. I don't know if you realized but in Critter1.2 it is mostly used just as a speed optimization. It is like lazy eval - it is wrong in some cases, but the gained speed makes up for it. Let me explain it briefly:
We tried the hash move - did not cut, we try good/even capures (if any) - they did not cut either, now in most cases there is almost zero chance that a quiet move will cause a cut-off. Despite that we try some quiet moves. How many? Depends on the remaining depth. Closer to leaves less, closer to root more. We do a sequence: "make_move", "evaluate the position", not good (evaluation did not rise)? - "undo move" and try the next. Without the "expected gain" we would do this for many quiet moves. This slows us down. Now with "expected gain" table we can spare these 3 operations because we can say with quite a high probability that the move will not rise the evaluation to the expected threshold (or above) and thus we will be faster. What if we are wrong? Bad luck. This is a gamble. But it is done more frequently at the leaves where errors are inherent anyway due to QS (=considering only captures). This is not worse than lazy eval, which must be wrong from time to time to be effective. [I understand that you had no luck with lazy eval either :(]

> Our lower-depth pruning is similar to Stockfish, not to Ippo


Pardon? They are very alike in this respect. It is just that SF gambles slightly more (and reaches a ply or two deeper thanks to that)...

> If you tuned LMR etc. yourself, why are the values you use so similar to or identical to Ippo in so many cases?


I do not care much about fine tuning of values. If I implement some idea, it either works or it doesn't. For the first implementation I use a formula/value that is at least in the same ballpark. If it does not help, I try to go one notch higher, then one notch lower, and if neither works I am tempted to scratch the idea altogether. But - just like You & Don - i try to understand why it is beneficial to program A and not for program B. If unsure I implant the thing to a "foreign" engine (I have an experimental StockFish1.9) and see if it works here. Etc.. Etc..
Back to your question - I use stuff which came out of my tests as best. I am not suffering with a not-invented-here syndrome, and I guess some specifics are just because of being state-of-art. Unless something better is invented _every_ top engine will employ same formulas with some variations. Think of LMR being a log2(n) formula

> My template comment was based partly on your own statement that your Pascal program was "heavily Ippo-ized" or something like that.


I made it public for 2 reasons:
a) it is of no use, it is slow like hell and it was just a testbed for ippo(+some houdini) stuff
b) nobody will care to clone it because it is written in Pascal (real men code in C) :)

> I ran some experiments together with Mark Leffler in which we tried turning off various ideas and formulas in one of the Ippo versions, and we were surprised to discover that in many cases the ideas or formulas were worthless or worse.


I don't know about Mark Leffler's understanding of a modern chess engine and how systematic your tests were. Just randomly switching stuff off without reason doesn't bode well... Nevertheless Ippolit isn't ideal both from algorithmic nor chess-knowledge point of view.

> This is probably why we have been able to pass Ippo despite a large speed handicap.


I have a rather different view on this, but let's leave this for another discussion...
Parent - - By lkaufman (*****) Date 2012-01-08 18:42
We do pretty much the same thing in terms of looking at just a few moves near the leaf nodes, but without the expected gains idea; we just rely on regular move ordering. We didn't find the expected gains method to be any better. Of course it's possible something wasn't implemented right, although Don was pretty sure that it was.
     Unless I'm mistaken, Stockfish prunes near the leaf nodes based on movecount alone (as do we), while Ippo, Houdini, and Critter use two different formulas based on both movecount and score. Again, we didn't find any benefit from the fancy formulas as opposed to just counting moves.
     I think you are wrong about Stockfish gambling more in low-depth pruning. They gamble LESS, using much higher move-count numbers than the Ippos or Critter, as well as higher thresholds for stuff life static null move. The Stockfish speed comes from aggressive (over aggressive I think) LMR in my opinion. That's why it doesn't show up at very low depths.
     As I understand it, the reason Critter formulas are often identical to Ippo is that you tried them, got a good result, and were not interested in tuning the values, correct? Nothing wrong with that unless you want to enter an ICGA tournament. We fine-tune everything so our values never look like anyone else's values, which in my opinion are all poorly tuned (for the programs whose values I know).
     The bottom line is that while you have done nothing wrong, Critter ends up behaving much like the Ippos, because you have found that their stuff works. Komodo is not like the Ippos because we have found that much of their stuff does not work (for us). I hope you would agree that this is a fair summary of the situation. I want to make it clear that this is just an observation, not a criticism of you or Critter, both of which I think very highly of.
     In my tests with Mark Leffler we didn't "randomly" turn off stuff; we mostly turned off stuff that was not helpful in Komodo, and in general found that these things were also not helpful in Ippo. That convinced me that the problem was not with Don's implementation but with Ippo itself.
     As for why we have passed Ippo, there are of course many reasons. Not using many Ippo ideas is just one factor.
Parent - - By Banned for Life (Gold) Date 2012-01-08 20:31
As I understand it, the reason Critter formulas are often identical to Ippo is that you tried them, got a good result, and were not interested in tuning the values, correct? Nothing wrong with that unless you want to enter an ICGA tournament. We fine-tune everything so our values never look like anyone else's values, which in my opinion are all poorly tuned (for the programs whose values I know).

This is very interesting. So if you want to enter an ICGA tournament, you have to tune your engine's chess knowledge so that it doesn't match the chess knowledge in another program? Suppose more programs were well tuned and ended up with very similar values. Would that make them all unoriginal? Or would later engines have to accept inferior weights? And how deep does this go? Can different programs use common piece values and combination bonuses, or do these need to be modified so that they never "look like anyone else's values". Isn't this a bit silly?
Parent - - By Richard Vida (**) Date 2012-01-09 01:28

> This is very interesting. So if you want to enter an ICGA tournament, you have to tune your engine's chess knowledge so that it doesn't match the chess knowledge in another program? Suppose more programs were well tuned and ended up with very similar values. Would that make them all unoriginal? Or would later engines have to accept inferior weights? And how deep does this go? Can different programs use common piece values and combination bonuses, or do these need to be modified so that they never "look like anyone else's values". Isn't this a bit silly?


I predict that Komodo _will win_ the next ICGA championship. Although I am not sure how much value is in competing against 2300 elo engines :(
Parent - - By Banned for Life (Gold) Date 2012-01-09 01:34
If this year was a guide, Komodo MP will be like a giant among pigmies in next years tournament...
Parent - By Arrière Pensée (Gold) Date 2012-01-09 02:10
Even pigmies can do learn to do the Komodo rag!
Parent - By Arrière Pensée (Gold) Date 2012-01-09 02:11
Even pigmies can do learn  the Komodo rag!
Parent - By lkaufman (*****) Date 2012-01-09 02:18
Well, I don't make the ICGA rules. It will always come down to a judgment call on whether a given engine is too similar to other ones. If your formulas, values, tables, etc. are so similar to another program that there is no chance that two independent programmers would come up with the same values then they should not both be allowed to play.
Parent - - By Richard Vida (**) Date 2012-01-09 01:20

> Unless I'm mistaken, Stockfish prunes near the leaf nodes based on movecount alone (as do we), while Ippo, Houdini, and Critter use two different formulas based on both movecount and score. Again, we didn't find any benefit from the fancy formulas as opposed to just counting moves.


more power to you...

> I think you are wrong about Stockfish gambling more in low-depth pruning. They gamble LESS, using much higher move-count numbers than the Ippos or Critter, as well as higher thresholds for stuff life static null move.


Then I am afraid you don't really understand how the search works in mentioned programs... At low depths SF is the _most_ selective engine I ever saw. Critter is much less selective.

> Komodo is not like the Ippos because we have found that much of their stuff does not work (for us).


Komodo is very Ippo like (minus the used weights/formulas). I think you would be very happy if you managed to get working the remaining two ideas from Ippo not already in K3. (namely - "lazy eval" & "expected positional gain")
Parent - By lkaufman (*****) Date 2012-01-09 02:39
By low depth let's use the last four plies, which I believe is the definition in Ivanhoe and corresponds to what you call "selective search". If you compare the pruning formulas such as movecount pruning, static null move, futility, etc. in each case Stockfish is more conservative than Ivanhoe or Critter AT THOSE DEPTHS. I can only surmise that you are using "low depth" to mean something like the last 7 or 8 plies. If I'm mistaken and you actually mean the last four plies, please name one or two ways in which SF is more selective on those plies. I'm pretty sure I'm right because SF is much slower than Ivanhoe or Critter in 7 ply search games. But if I am somehow wrong please educate me.
     When you say that Komodo is "Ippo-like", do you mean that just in the sense that Stockfish is also "Ippo-like", or do you think Komodo has much more in common with Ippo than SF does? Offhand, I can think of only one pruning formula that would make Komodo any closer to Ippo than Stockfish is, and as you said SF uses expected gain and Komodo does not.
     You are right that if we got a good gain out of expected gain and/or lazy eval we would be quite happy, that would be enough to catch Houdini at least at the longer time controls. But we also don't use the Ippo (really Rybka) ideas of pruning at low depth based on score combined with movecount and of course lots of little things.
1 2 3 4 5 6 7 8 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill