I told you that sometimes you do make good suggestions, and that is good, but sometimes you don't, but that is not the problem. Everyone has some ideas that work and some that don't.Lyudmil Tsvetkov wrote:This time you already insult me, Arjun, but I take no offence.arjuntemurnikar wrote:It is sad that all that experience gave you no sanity.Lyudmil Tsvetkov wrote:There are lots of Buddhas around this forum, Arjun.arjuntemurnikar wrote:This is an insult to Joona and all his efforts in tuning stockfish long before fishtest. SF is 3000+ elo now because of his tuning efforts. If it doesn't work -- either it is already well tuned, or the setup is wrong. And if you ask me, yes I have had success using CLOP and SPSA in my private testing. You on the other hand, have never used it, nor have you even the slightest clue how it works, but you see a RED color in fishtest, and you immediately jump to your "expert" conclusions!Lyudmil Tsvetkov wrote:It is the first time that someone accuses me of being disrespectful of people here. I do not remember an occasion where I have insulted someone. But I stick to what I believe of course.arjuntemurnikar wrote:An intuitive approach is hardly sensible!Lyudmil Tsvetkov wrote:Hi Uri.Uri Blass wrote:I think that maybe reducing the middle game value of the knight value by 2% may give a better result if the result of big elo loss of 4 rating points hold
with more games.
stockfish see advantage for black in the following position from my game that is not correct
r1bq1rk1/p3nppp/1pn1p3/3pP3/1P1N4/P1B5/2P2PPP/R2QKB1R w KQ - 1 11
I saw so many games where SF underestimates knights in the mg preferring instead bishop, pair of bishops, pawns, queen, etc. A really astounding number of games where that would be true.
And yet the increase seems to fail currently convincingly. Some things are really difficult for me to explain in SF. I am sure a lot could be gained by tweaking piece values in SF, but, if a sensible approach does not work, I do not know what will.
What is sensible is employing a mathematically-proven algorithm to tune the values in a systematic way.
I already explained this to you Lyudmil, but you will never listen to people even when it is regarding a subject which you know absolutely nothing about (SPSA and CLOP). Your utter disregard and disrespect for people who give you good and patient advice is what annoys me most.
As for the matter of piece values, I believe this has been tuned to death in the past by Joona, and there is hardly anything left there, but maybe with 1,000,000 SPSA games, we may gain 1 elo or something... (Hardly something a sane person will pursue as potential for elo!), and even if there is more elo, the way to find out is through algorithmic tuning, not "random fantasy value insertions".
How many successful patches you had with SPSA?
There are other factors in eval that contribute to the apparent reduction and increase in piece values in different positions. So, completely wrong. If the piece values were so badly calibrated, SF would not be 3000+ right now. If there is any tuning left, it is few elo at most!Lyudmil Tsvetkov wrote: Piece values in SF are not tuned well at all. Actually, currently they are incredibly inaccurate. For example, you have knight value increasing from mg to eg with some 30cps, and bishop value with just 20cps, when everyone knows that the bishop gains more strength relative to the knight in the eg. Similarly, rook value increases from mg to eg with just 8cps, when the rook in general is supposed to have a much higher value in the eg. Those are not well-tuned values. Rather, they are the product of tuning in a larger term pool involving redundancies here and there.
... and now you blame the search, which has absolutely nothing to do with this matter! When you don't understand something, blame the magic unicorn. Sure!Lyudmil Tsvetkov wrote: However, this is not the biggest problem for me. I am almost certain there is something specific into the search that would prevent to correct some parameters, even if this is the right approach. The old parameters, even if basically wrong, will resist alteration because functioning in tune with some specific search parameters. That is how I explain the failed knight test.
"Every second game" is an exaggeration, but even then, I think here your mind is biasing you to see a pattern where there is none. You are so hard-bent on proving piece values are untuned that every other game you see, your mind leads you to the same false conclusion. There is a psychological term for this, but I forget what it is called.Lyudmil Tsvetkov wrote: Considering the knight value, how can you say that this is a stupid intuitive approach, when in every second game involving imbalance I have looked at that SF lost, SF always underestimated knights and that was the reason for losing. Do you understand, every second game. I think this is more a mathematically sound approach, not an intuition. It simply is impossible not to work, but still it does not. So there is something into the search. I do not understand, people could help here, but I am certain it has to do with the search. What search features could prevent sensible tuning?
So again, here is an example: you look at 100 games (all available SF lost games), and in 50 of them you observe that SF loses because it evaluate positions with knights worse than positions with other pieces and pawns. What is the logical, mathematically-based conclusion? That you should increase the knight value. You do that, and it does not work. How could you explain that?
Anyway, I have no idea how given just
1. The position
2. The final evaluation score
...you can come to the conclusion that in the given position SF over-values/under-values a knight! You must be GOD or something!!
There are 100+ eval features that are computed in a not-so-straightforward way by SF, so how can you breakdown exactly what is happening by just looking at the position and the final score??
If you ran eval command, and then you made a few claims, then I would take you seriously, but you don't. You make claims out of thin air.
To make it clearer (because I know you don't understand), how do you know in a particular position SF is over-valuing the knight, when there are other features of the position like pawn-structure, king-safety, etc. that may be counter-acting the knight evaluation? Without running eval and without proper debugging, any claim, especially about code which has been already well-tuned, is just silly.
I am not trying to demotivate you. Suggestions are always good (and you do make very good suggestions now and then). It is just that, when most people make bad suggestions, and somebody explains to them nicely why they are bad, then that's the end of it. Everybody lives in peace. But when YOU make a bad suggestion, and somebody explains to you nicely why it is bad, you ignore the person, question his merits, and continue to claim that you are right!
If someone does not get annoyed by this, they must be Buddha or something.
Not 1 or 5 elo, I am absolutely certain SF can gain 50 elo just on better tuned piece values. Absolutely certain. So SF 6 could be just a change of piece values.
When I say I have looked at 100 games, this means I have looked at 100 games, and when I say that in half of the games SF underestimates knights, that means it does so in half of the games. I have no reason to lie.
I do not need to run eval command, besides it shows just a couple of terms. For an experienced chess player, accustomed to SF playing style, it is obvious what the engine is doing. There are positions where king safety is irrelevant, or pawn structure is irrelevant, etc. That is all the trick: to quickly be able to see what happens on the board. The quicker you get it, the more experienced you are. It is like that in every area. Experience comes with practice, that is it.
SF performing at 3000 does not mean that it has no glitches, but rather than other engines perform at lower elo. While SF piece values are badly tuned (Marco does not get offense here, you see:D), other engine piece values are even worse. Very simple.
I tell you again: it is not logical that the knight value increases from mg to eg more than the bishop value, very simple. And therefore SF has problems with its knights in games. It is not logical that the rook value stays almost the same in mg and eg, and therefore SF underestimates rooks in eg. Etc. You had a chance that in TCEC SF was the much better opponent overall, otherwise you would have seen what big problems it has with its piece values and imbalances! It is visible even now, but much weaker Komodo was unable to convert. Look at the last season, SF lost so many games because of its piece values. Its eval did not improve in that area substantially since then, so the problems have not disappeared. If SF misplayed 1/3 of the positions in the previous final based on its deficient piece value/imbalance understanding, then there is a problem, and it is huge. In the last TCEC SF simply had too much other advantages for this to show clearly. But it will certainly show with stronger opponents.
So that I stay by that: SF 6 could be just a change of piece values.
You can continue ignoring good advice, and you can continue re-iterating your same points over and over again. Sometimes innocent people will come around and will push a random test or two for you around piece values. They will fail. But, you will continue to claim you are right, and continue to claim that somehow magically, SF search is messing with your exact values.
I have come to know you very well now, Lyudmil, and I think sometimes you are just full of crap.
My ideas are bad, but a storming pawn idea worked, a low mobility idea worked, a chain rank bonus idea worked, a queen imbalance idea worked, an advanced levers idea worked.
How many of your creative/non-simplification ideas actually worked?
The problem is that when nice people explain to you in nice terms why an idea is bad, you ignore them, question their merit and continue making claims that you are actually correct. And worse, it is often in a subject that you know less about, and the person explaining to you knows what he is talking about. And even WORSE, when one of those bad ideas is tested in fishtest, and it fails, you blame another part of the code for interfering with your "perfect intuition".
Either you have a big ego, or no sense at all.
I am sorry you feel insulted, but you insult me first by ignoring everything I said. I put time and effort typing those words for you, but they go through one ear and come out of your other like air...
