bob wrote: I can answer that. zero. From testing. I removed them from Crafty and found Elo went up. I removed them from SF 1.6, one by one, and found that the elo either went up slightly or did not change. Except for the check extension itself which is worthwhile although not +20 worthwhile...
Ok, I will remove all the extension at once, apart from SE and check exstension, and test and if there is no advantage we will remove all them in one go
That's an OK test. I think it better to go one at a time, because each chess engine is different, with different rules for things like reductions, pruning, etc.
You said that you already tried to remove one by one in SF 1.6, so if removing all does not have any effect then I will trust you and take as valid your results
Should work. I was making a general comment that what works for one doesn't always work for another. Anyone following the "extension thread" ought to test their own engine carefully rather than just blindly playing "follow the leader" like lemmings.
Was somewhat disappointing for me since I had been doing all of these extensions for many years, dating back to Cray Blitz which even did the full Hsu SE. I still plan on testing that, but it is quite complex if you read the description. I now wonder if that helped Cray Blitz at all (or even deep thought for that matter). It might have been more valuable back when we were seeing 8-10 ply search depths in 3 minutes.
Bob, do you think at some stage even the good-old check extension won't be worth doing? I mean in 10 years, assuming we keep near to Moore's law, we will be able to search loads deeper. At some stage, I guess it won't be even worth extending checking moves..
This might well be true. The check evasion extension (today) is not a huge gain. I no longer have the exact numbers but will try to run current 23.4 with check extension disabled and will post the results...
That will be interesting.. my guess is that it's worth a lot less now than when we were all doing 6 ply searches. Maybe it's somewhere as little as 20ELO?
If my memory is right, it is somewhere at _most_ +20 Elo.
bob wrote: I can answer that. zero. From testing. I removed them from Crafty and found Elo went up. I removed them from SF 1.6, one by one, and found that the elo either went up slightly or did not change. Except for the check extension itself which is worthwhile although not +20 worthwhile...
Ok, I will remove all the extension at once, apart from SE and check exstension, and test and if there is no advantage we will remove all them in one go
After 5404 games
No_extension vs Original version: 872 - 963 - 3569 ELO -5 (+- 3.8)
I don't have enough resolution to test one by one, so I think extension will still stay there.
People please, if you can, avoid stuff like "you should test with many engines" and the usual boring advices (they are not requested). Thanks in advance.
I am running this again, first with Crafty. I am afraid, however, that my test might be just a bit broken but am going to wait. If I have "extensions" > 0, which is true when doing a check, I do not reduct. By setting the check extension to zero, I do reduce. So this will be a lower bound on the Elo with no check extension, but if it is more than 15 Elo down, I will change the way I do this so that I do not reduce checks, which is probably better. More as the test results come in, I have two 23.4 runs queued up first to verify recent changes...
bob wrote: I can answer that. zero. From testing. I removed them from Crafty and found Elo went up. I removed them from SF 1.6, one by one, and found that the elo either went up slightly or did not change. Except for the check extension itself which is worthwhile although not +20 worthwhile...
Ok, I will remove all the extension at once, apart from SE and check exstension, and test and if there is no advantage we will remove all them in one go
That's an OK test. I think it better to go one at a time, because each chess engine is different, with different rules for things like reductions, pruning, etc.
You said that you already tried to remove one by one in SF 1.6, so if removing all does not have any effect then I will trust you and take as valid your results
Should work. I was making a general comment that what works for one doesn't always work for another. Anyone following the "extension thread" ought to test their own engine carefully rather than just blindly playing "follow the leader" like lemmings.
Was somewhat disappointing for me since I had been doing all of these extensions for many years, dating back to Cray Blitz which even did the full Hsu SE. I still plan on testing that, but it is quite complex if you read the description. I now wonder if that helped Cray Blitz at all (or even deep thought for that matter). It might have been more valuable back when we were seeing 8-10 ply search depths in 3 minutes.
Bob, do you think at some stage even the good-old check extension won't be worth doing? I mean in 10 years, assuming we keep near to Moore's law, we will be able to search loads deeper. At some stage, I guess it won't be even worth extending checking moves..
This might well be true. The check evasion extension (today) is not a huge gain. I no longer have the exact numbers but will try to run current 23.4 with check extension disabled and will post the results...
That will be interesting.. my guess is that it's worth a lot less now than when we were all doing 6 ply searches. Maybe it's somewhere as little as 20ELO?
I do not know but my guess is that maybe heavy pruning reduce the value of the extensions and the value of check extensions is bigger without LMR that is not used for checking moves(with LMR even without check extensions the fact that programs do not reduce checking moves in LMR already means that the program search checks deeper relative to most moves).
bob wrote: I can answer that. zero. From testing. I removed them from Crafty and found Elo went up. I removed them from SF 1.6, one by one, and found that the elo either went up slightly or did not change. Except for the check extension itself which is worthwhile although not +20 worthwhile...
Ok, I will remove all the extension at once, apart from SE and check exstension, and test and if there is no advantage we will remove all them in one go
After 5404 games
No_extension vs Original version: 872 - 963 - 3569 ELO -5 (+- 3.8)
I don't have enough resolution to test one by one, so I think extension will
still stay there.
People please, if you can, avoid stuff like "you should test with many engines" and the usual boring advices (they are not requested). Thanks in advance.
What about overlap between reductions and extensions. Are there positions where SF extend and reduce at the same time?
bob wrote: I can answer that. zero. From testing. I removed them from Crafty and found Elo went up. I removed them from SF 1.6, one by one, and found that the elo either went up slightly or did not change. Except for the check extension itself which is worthwhile although not +20 worthwhile...
Ok, I will remove all the extension at once, apart from SE and check exstension, and test and if there is no advantage we will remove all them in one go
After 5404 games
No_extension vs Original version: 872 - 963 - 3569 ELO -5 (+- 3.8)
I don't have enough resolution to test one by one, so I think extension will
still stay there.
People please, if you can, avoid stuff like "you should test with many engines" and the usual boring advices (they are not requested). Thanks in advance.
What about overlap between reductions and extensions. Are there positions where SF extend and reduce at the same time?
For the record, the check extension removal (done correctly this time) is a -60 Elo change. This version doesn't extend checks, but also does not reduce them unless they are unsafe, just like the normal check extension version. This -60 is probably on the high end for most, as I got 6-8 Elo just by stopping the extension if the move is unsafe according to SEE (< 0).
For the record, the check extension removal (done correctly this time) is a -60 Elo change. This version doesn't extend checks, but also does not reduce them unless they are unsafe, just like the normal check extension version. This -60 is probably on the high end for most, as I got 6-8 Elo just by stopping the extension if the move is unsafe according to SEE (< 0).
Ralph Stoesser wrote: What about overlap between reductions and extensions. Are there positions where SF extend and reduce at the same time?
That would seem (to me) to be insane.
An extendible move raises 'dangerous' flag that inhibits LMR and pruning on that move.
Same as what we do. Would make little sense to extend by 1, then reduce by one, for a net change of zero. Those two actions should be mutually exclusive. If a move deserves an extension, it can't deserve a reduction, and vice-versa...
