mcostalba wrote:Lyudmil Tsvetkov wrote: Again, for those who doubt it: specific knowledge works, general knowledge often fails.
I agree, but in the seldom case when general knowledge succeeds, it replaces a lot of specific knowledge and you end up with a better (in broader sense) engine, not only a stronger one...and when you have a better engine it is easier to further add another step of improvement, and this is really the key why you want a 'better engine' an not only a stronger one.
In some way it is like to add little cubic bricks one above the other, you can reach higher heights adding on top of previous one brick at a time...but sooner or later you'll get a very tall and narrow tower when adding more bricks becomes difficult and with risk of imploding....only when you add the bricks in a way to expand the tower basement, you perhaps don't make it _immediately_ higher, but pave teh way for further improvements and you end up building a pyramid that, as we know, last much longer...and can be built to be even taller!
Thanks Marco, for sharing.
I perfectly agree with this, absolutely. An engine with less but more important and better tuned eval terms should work better. There are terms you could replace and terms you can not. For example, you can not replace mobility, piece king attacks, some special imbalances, etc. The terms that are irreplaceable should be there, other you can remove.
I tell you I am very happy SF tried some of my suggestions, really very happy. But you know, people tried only the least valuable and most replaceable ones.
The things that easily fitted to SF code. Like some mobility tweaks, other tweaks, etc. Those terms are still replaceable. However, you can not replace terms like play in closed positions, blockade, special material imbalances, some other specific knowledge, etc.
My only concern is SF elo, when will SF add another 20 elo to SF 5? This could be the longest-running release. And of course, you know, I again see just 1 test in STF, it is sorrowful.
I tell you again some very important eval code SF needs from my point of view:
- better space eval, maybe try to improve on the old one
- insufficient imbalance eval, still many things to do there
- improved storm pawn code
- scaling down of opposite colour bishops, absolutely necessary for SF not to look funny. I think this will bring also nice elo
- implementation of blockade and closed positions eval
- bigger penalty for doubled and isolated pawns (you know, there are tens of games where SF fails on this, why not try to improve it), etc., etc.
I fully agree it is not the best way to implement separately low bishop, low queen, low knight, etc. mobility. Gary did a great thing, with a patch handling all those issues as one, I think this is the way to go. Also, concerning imbalances, it is not te best way to handle all imbalances separately, but why not have a single patch where Q vs 3 minors, Q vs 2 minors and R, Q vs 2Rs and minor, R vs NBB and BNN, 2Rs vs 3 minor pieces are specified with their respective different values. As those patches score positively separately, they should make it in a unified patch and the engine play better consequently. That knowledge is irreplaceable. And it should not slow down the engine, as it is almost the same as specifying different values for pawns on different files and ranks.
Sorry Marco, I again drifted off. But SF really needs good structured new knowledge. When will SF 5 be released, when?....
