odyn1982 wrote: ↑Thu Feb 20, 2020 11:23 pm
I want to use this engine mainly as a game partner because of some great features: personalities, Auto Skill, ELO adjustment.
In that case you wouldn't be able to tell the difference of 15 elo with the update.
Furthermore, this 15 ELO is measured against computers, not humans, it's possible you already have the strongest version against humans (for instance, increasing Contempt beyond a certain point provides a much weaker ELO engine, but one much stronger against humans.)
odyn1982 wrote: ↑Thu Feb 20, 2020 11:23 pm
I want to use this engine mainly as a game partner because of some great features: personalities, Auto Skill, ELO adjustment.
In that case you wouldn't be able to tell the difference of 15 elo with the update.
Furthermore, this 15 ELO is measured against computers, not humans, it's possible you already have the strongest version against humans (for instance, increasing Contempt beyond a certain point provides a much weaker ELO engine, but one much stronger against humans.)
I'm quite convinced that Komodo MCTS is stronger against humans than standard Komodo, even though it is weaker by perhaps 70 elo against other engines and by even more in direct play. This is clear from the handicap matches with masters and grandmasters, and also clear to me when I play against it at knight odds. The MCTS version is just more aggressive and tricky, and doesn't just try to maximize how many moves it will take to lose. Maybe in even game play normal Komodo might score higher against me, perhaps 99% instead of 98% if they both use Contempt of 100 or so, but even that is questionable since the trickier MCTS might lead to fewer draws. I think that Komodo MCTS is the best engine in the world now at beating strong human players with some handicap that gives them a chance.
Regarding the extra features added since 13.1, that is indeed a valid reason for an upgrade for the Android. Most likely it would happen after we release Komodo 14, which based on past years would be in May (no promise!).
I'm quite convinced that Komodo MCTS is stronger against humans than standard Komodo, even though it is weaker by perhaps 70 elo against other engines and by even more in direct play
I am quite convinced that your blended MCTS/AB approach is light years ahead of the grossly inadequate search of LC0. That is why I recently renewed my subscription to help support your efforts. I hope at some point you try applying your ideas to a NN running on GPU hardware. I think you and Mark are on the right path to improving the horrible search design of MCTS alone. I often wonder if you could do MCTS using a NN with AB rollouts using your human built CPU friendly evaluation function.
Best of luck and thank your for decades of top ideas,
Mike
“Censorship is telling a man he can't have a steak just because a baby can't chew it.” ― Mark Twain
Krzysztof Grzelak wrote: ↑Fri Feb 21, 2020 8:04 am
I understand that when you buy a Komodo chess engine there is no version MCTS
MCTS is a UCI option - its off by default, turn it on with a checkmark. Typically you would also set multipv to 4 or so for best analysis. So one exe - but two engines.
I'm quite convinced that Komodo MCTS is stronger against humans than standard Komodo, even though it is weaker by perhaps 70 elo against other engines and by even more in direct play
I am quite convinced that your blended MCTS/AB approach is light years ahead of the grossly inadequate search of LC0. That is why I recently renewed my subscription to help support your efforts. I hope at some point you try applying your ideas to a NN running on GPU hardware. I think you and Mark are on the right path to improving the horrible search design of MCTS alone. I often wonder if you could do MCTS using a NN with AB rollouts using your human built CPU friendly evaluation function.
Best of luck and thank your for decades of top ideas,
Mike
Well, we have a version of Komodo MCTS that uses a GPU with an NN blended with Komodo MCTS, which we have used in some knight odds matches with human masters with great results. The problem is that even with a good GPU, it is only about even with standard Komodo, and we don't think that many people would be keen to buy a Komodo version that needs an expensive GPU if it isn't clearly stronger than standard Komodo. Once we have a solution to this problem, we plan to offer it to the public.