Alayan wrote: ↑Tue Nov 03, 2020 7:21 pm
This is not relevant to what I said.
My point was not about bad search + good eval vs bad eval + good search. Yes, bad eval + good search wins this one, I agree.
My point was and is that K eval + SF search wouldn't crush SF eval + K search (abstracting away some minor search/eval compatibility issues coming from different eval scaling and such), and would maybe even lose to it.
I think the other half of the story that many people missed was that when Leela came on the scene two and a half years ago and started finding holes in all the other engines' evaluation functions, the Stockfish developers started implementing ideas from Leela's play like thorn pawns that boosted the strength of Stockfish's evaluation function and lead to better positional play from Stockfish.
On the other hand, Komodo never took inspiration from Leela for improving its evaluation function, and instead went on a huge detour with MCTS search which has only been proven to be good for MultiPV analysis. And because of that, Komodo's evaluation function stagnated, and Stockfish's evaluation function was able to surpass Komodo's.
Last edited by Madeleine Birchfield on Tue Nov 03, 2020 9:05 pm, edited 1 time in total.
Alayan wrote: ↑Tue Nov 03, 2020 7:21 pm
This is not relevant to what I said.
My point was not about bad search + good eval vs bad eval + good search. Yes, bad eval + good search wins this one, I agree.
My point was and is that K eval + SF search wouldn't crush SF eval + K search (abstracting away some minor search/eval compatibility issues coming from different eval scaling and such), and would maybe even lose to it.
I think the other half of the story that many people missed was that when Leela came on the scene two and a half years ago and started finding holes in all the other engines' evaluation functions, the Stockfish developers started implementing ideas from Leela's play like thorn pawns that boosted the strength of Stockfish's evaluation function and lead to better positional play from Stockfish.
Leela's presence created more competition and so more motivation to work on Stockfish. But the influence that Leela had over specific elo-gaining eval patches to try and make SF play more like Leela is vastly overblown, it was a handful at best.
Alayan wrote: ↑Tue Nov 03, 2020 9:08 pm
Leela's presence created more competition and so more motivation to work on Stockfish. But the influence that Leela had over specific elo-gaining eval patches to try and make SF play more like Leela is vastly overblown, it was a handful at best.
Whatever the case, what may have been true in 2016 and 2017 (Komodo evaluation being stronger than Stockfish evaluation) is obviously no longer true in 2020.
2500 at knight handicap ? Woo. How much stronger it is than Komodo MTC against human in this handicap ? 100 elo perhaps ? You finally have the best program ever for this handicap.
lkaufman wrote: ↑Tue Nov 03, 2020 8:19 pm
Chess.com CEO thinks it's important to have the Dragon name first, so that people will call it that, so for rating lists and such we suggest listing it as Dragon (Komodo). But of course some may call it Komodo Dragon, just as people's names can be first or last name first.
I think to have the Dragon name first can lead to confusion. Because there is already a chess program named Dragon.
If they are very keen on a dragon name, why not Wyvern or Smaug. I don't think either of those are used yet.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
