SzG wrote:I have always wondered why Rybka cannot promote to bishop (while it accepts a bishop promotion). Is checking for this move so time-consuming that it isn't worth it? Does it make the code much more complex? Then why do all other engines do it?
Can it be that this thing has some big secret in it, maybe the reason of Rybka's great strength? Vasik has been attacked many times for the lack of bishop underpromotion but he stubbornly declined to implement it. But why?
It saves a minor amount of time, so it helps Rybka's strength slightly since it is incredibly rare in real games to be helpful (pretty much only queens and knights are relevant in non-composed problems). I did the same thing in my engine. It is not a big enough deal to explain Rybka's strength, just one of many things that are slightly helpful.
SzG wrote:I have always wondered why Rybka cannot promote to bishop (while it accepts a bishop promotion). Is checking for this move so time-consuming that it isn't worth it? Does it make the code much more complex? Then why do all other engines do it?
Can it be that this thing has some big secret in it, maybe the reason of Rybka's great strength? Vasik has been attacked many times for the lack of bishop underpromotion but he stubbornly declined to implement it. But why?
An engine that is 100% faster is 50 Elo stronger.
An engine that is 10% faster is 5 Elo stronger.
Skipping bishop promotions makes Rybka 10% faster and gains 5 Elo points.
Matthias Gemuh wrote:An engine that is 100% faster is 50 Elo stronger.
An engine that is 10% faster is 5 Elo stronger.
Skipping bishop promotions makes Rybka 10% faster and gains 5 Elo points.
You are on the low side. Many (educated) people would say:
An engine that is 100% faster is 70 Elo stronger.
An engine that is 10% faster is 10 Elo stronger.
"Skipping bishop promotions makes Rybka 10% faster"
Where does that come from?
ernest wrote:
"Skipping bishop promotions makes Rybka 10% faster"
Where does that come from?
I think it was a hypothetical. Skipping some under-promotions is not a new idea, and many authors have tried it out. If you are really interested, I am sure Bob has collected stats on exactly how much it effects crafty (I am confident it helps very slightly and likely he keeps under-promotion for reasons of principal not related to engine strength).
Matthias Gemuh wrote:An engine that is 100% faster is 50 Elo stronger.
An engine that is 10% faster is 5 Elo stronger.
Skipping bishop promotions makes Rybka 10% faster and gains 5 Elo points.
You are on the low side. Many (educated) people would say:
An engine that is 100% faster is 70 Elo stronger.
An engine that is 10% faster is 10 Elo stronger.
"Skipping bishop promotions makes Rybka 10% faster"
Where does that come from?
"makes Rybka 10% faster" is more readily understood by non-programmers than "reduces Rybka's search tree by 10%".
I think Vas could easily make an "analysis" engine for such problems where losing some speed will not be a problem. But for real games it is not an issue. The bishop under promotion might happen once in every 2 million games, and in those 2 million games 99% would have been decided earlier than the point of bishop under promotion ... so I find it hard to use the term "cripple". More than likely the extra 5 or 10 ELO that the engine gains will be several tens of thousands of games won during the same number of games.
I think we might see some different personalities in Rybka 4 as this is very easy for Vas to do. Adding knowledge code slows down an engine dramatically, and in many cases "dumbing" down an engine by removing code makes it play stronger in fast time control games simply because it calculates faster. Eventually this catches up with you in longer time controls. Rybka Winfinder was like that, it had a lot of knowledge removed but could find tactical shots sometimes 10X faster than normal Rybka.
SzG wrote:I have always wondered why Rybka cannot promote to bishop (while it accepts a bishop promotion). Is checking for this move so time-consuming that it isn't worth it? Does it make the code much more complex? Then why do all other engines do it?
Can it be that this thing has some big secret in it, maybe the reason of Rybka's great strength? Vasik has been attacked many times for the lack of bishop underpromotion but he stubbornly declined to implement it. But why?
An engine that is 100% faster is 50 Elo stronger.
An engine that is 10% faster is 5 Elo stronger.
Skipping bishop promotions makes Rybka 10% faster and gains 5 Elo points.
Matthias.
I don't believe it is worth 10%, I'll measure this later this afternoon and report actual results.
ernest wrote:
"Skipping bishop promotions makes Rybka 10% faster"
Where does that come from?
I think it was a hypothetical. Skipping some under-promotions is not a new idea, and many authors have tried it out. If you are really interested, I am sure Bob has collected stats on exactly how much it effects crafty (I am confident it helps very slightly and likely he keeps under-promotion for reasons of principal not related to engine strength).
-Sam
It is trivial to turn off in Crafty (there is a simple place where the under-promotions are generated sequentially). I found zero difference when I ran this test on the cluster the last time this came up. I also do not remember any measurable speed difference, but am going to re-run this later today to confirm. Can't do any cluster testing until the A/C issues are solved next week, however...
M ANSARI wrote:I think Vas could easily make an "analysis" engine for such problems where losing some speed will not be a problem. But for real games it is not an issue. The bishop under promotion might happen once in every 2 million games, and in those 2 million games 99% would have been decided earlier than the point of bishop under promotion ... so I find it hard to use the term "cripple". More than likely the extra 5 or 10 ELO that the engine gains will be several tens of thousands of games won during the same number of games.
The only reason I can imagine that Rybka has this problem, is based on Vas being too busy. Vas may have thought this was a good idea, did not test it, and later, he figured it was not worth to come back and fix it. Having separate engines for "analysis" and "game" may make sense for other parameters (null move verification for instance), not for this trivial fix. When do you try B promotion in search? when Q promotion did not work, so, that will be a position that is going to be pruned in the next ply most likely anyway. Is Rybka using LMR? then underpromotions should be move to the end of the list. Underpromotions should never be pruned close the root or on the PV, IMHO.
Miguel
I think we might see some different personalities in Rybka 4 as this is very easy for Vas to do. Adding knowledge code slows down an engine dramatically, and in many cases "dumbing" down an engine by removing code makes it play stronger in fast time control games simply because it calculates faster. Eventually this catches up with you in longer time controls. Rybka Winfinder was like that, it had a lot of knowledge removed but could find tactical shots sometimes 10X faster than normal Rybka.
