How much stronger is Komodo now, compared to its TCEC season 9 version?
(If it adds on another 150 elo, then i think this may be enough to win Season 10. like +50 to bridge the gap of then, +50 to equal sf's own upcoming improvements, and +50 to get far ahead and be undisputed greatest, together that would be +150 elo, which is needed and wonderful).
How i s Komodo getting on? (now, end of December 2016)
Moderators: hgm, Rebel, chrisw
-
- Posts: 8514
- Joined: Thu Mar 09, 2006 3:25 am
- Location: Jerusalem Israel
-
- Posts: 1167
- Joined: Thu Dec 25, 2008 9:07 pm
- Full name: Herbert L
Re: How i s Komodo getting on? (now, end of December 2016)
+150ELO - You are a bit overoptimistic.
-
- Posts: 1494
- Joined: Thu Mar 30, 2006 2:08 pm
Re: How i s Komodo getting on? (now, end of December 2016)
We have had some nice gains from the last release (Komodo 10.3).
At the long time controls played for TCEC, elo gains are typically less than that rating groups will show at the much shorter/fewer CPU games they can afford to run. Longer games means more draws. More cores also mean more draws (depending on scaling). So a 150 elo gain in a few months is not (at present) possible. Even developments of roughly a year's worth are not going to add up to 150 elo for programs near the top. Squeezing even a few elo is not easy. For example. Stockfish 8 is about 61 elo stronger than Stockfish 7 (CCRL 40/40 4 cpu). Komodo 10.2 shows on the same list as 48 elo stronger than Komodo 9 (CCRL 40/40 4 cpu). We are not done with the Komodo 10 series and you can expect a significant elo increase when we release the next version, but not 150 elo!
We do expect a scaling with more processors improvement over earlier Komodo versions, but we have not had enough machine time to characterize it for more than 4-8 threads/cpus.
At the long time controls played for TCEC, elo gains are typically less than that rating groups will show at the much shorter/fewer CPU games they can afford to run. Longer games means more draws. More cores also mean more draws (depending on scaling). So a 150 elo gain in a few months is not (at present) possible. Even developments of roughly a year's worth are not going to add up to 150 elo for programs near the top. Squeezing even a few elo is not easy. For example. Stockfish 8 is about 61 elo stronger than Stockfish 7 (CCRL 40/40 4 cpu). Komodo 10.2 shows on the same list as 48 elo stronger than Komodo 9 (CCRL 40/40 4 cpu). We are not done with the Komodo 10 series and you can expect a significant elo increase when we release the next version, but not 150 elo!
We do expect a scaling with more processors improvement over earlier Komodo versions, but we have not had enough machine time to characterize it for more than 4-8 threads/cpus.
-
- Posts: 8514
- Joined: Thu Mar 09, 2006 3:25 am
- Location: Jerusalem Israel
Re: How i s Komodo getting on? (now, end of December 2016)
Does it also mean more draws vs engines which it may otherwise win?
-
- Posts: 5960
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
Re: How i s Komodo getting on? (now, end of December 2016)
Longer time limits mean less mistakes so less chances for the better engine to win. This would not be true if the opening book had many positions that were already winning, but I don't think anyone uses a book like that to test. If they did we might see larger elo gains at longer time controls.S.Taylor wrote:Does it also mean more draws vs engines which it may otherwise win?
I can say that we have yet to observe diminishing returns (in terms of elo per month) with Komodo, which is quite surprising. Of course Stockfish can also say that.
Komodo rules!
-
- Posts: 47
- Joined: Mon Mar 07, 2016 12:46 am
Re: How i s Komodo getting on? (now, end of December 2016)
I am impatiently waiting for my Christmas present, HoHoHo
-
- Posts: 313
- Joined: Sat Apr 05, 2014 12:09 am
- Location: Neenah, WI, United States
Re: How i s Komodo getting on? (now, end of December 2016)
Didn't they just release K 10.2 in late october? Do you think almost two months is enough time for a good improvement? Did Mark or Larry say they will release a version around christmas? That would be great of course, but I didn't think they would release after two months of testing.Tdunbug wrote:I am impatiently waiting for my Christmas present, HoHoHo
Sincerely,
Tim.
-
- Posts: 12038
- Joined: Mon Jul 07, 2008 10:50 pm
Re: How i s Komodo getting on? (now, end of December 2016)
do you have any estimate how many elo you gain when you double the time from tcec time controls. I think people used to say about 80 elo. what is it now.?lkaufman wrote: I can say that we have yet to observe diminishing returns (in terms of elo per month) with Komodo, which is quite surprising. Of course Stockfish can also say that.
-
- Posts: 5960
- Joined: Sun Jan 10, 2010 6:15 am
- Location: Maryland USA
Re: How i s Komodo getting on? (now, end of December 2016)
80 elo for a doubling is an estimate for blitz levels; at TCEC levels it might be more like 40, perhaps even less if you assume many cores running. Others can give you more accurate numbers.duncan wrote:do you have any estimate how many elo you gain when you double the time from tcec time controls. I think people used to say about 80 elo. what is it now.?lkaufman wrote: I can say that we have yet to observe diminishing returns (in terms of elo per month) with Komodo, which is quite surprising. Of course Stockfish can also say that.
Komodo rules!
-
- Posts: 8514
- Joined: Thu Mar 09, 2006 3:25 am
- Location: Jerusalem Israel
Re: How i s Komodo getting on? (now, end of December 2016)
I hope this means that it will not lose ever again to Stockfish, not sf 8 and not the next one sf 9 either.lkaufman wrote:Longer time limits mean less mistakes so less chances for the better engine to win.S.Taylor wrote:Does it also mean more draws vs engines which it may otherwise win?
At the same time, it should know how to take advantage of any suboptimal moves of SF if there ARE any.
It should also have the board presence as good as Carlsen has! (to know the best ways to probe the other sides weaknesses).