Dann Corbit wrote: ↑Fri May 29, 2020 11:45 pm
I don't think the weak 6 ply searches matter at all.
You can get to 15 plies in an eyeblink.
If you like ultra-hyper-bullet game in one tenth of a second, then use Komodo.
I am not sure how the NN programs would do at ludicrously short time control.
It would be interesting to see an LC0/Komodo match at very high speed.
For reasonable time control on both play and analysis, Stockfish does very well indeed.
You are wrong that Komodo or Lc0 are better in ultra-fast games. At fixed time, SF_dev is by far the strongest engine on a regular PC in hyper-ultra-fast games (say 1s+0.01s), it is Lc0 surpassing it at slower games and Komodo approaching SF level at slower games. Simply, depth 6 is completed by SF in maybe a millisecond, practically unplayable times using time controls. But I was curious how the notion of "ply" has changed over time comparing a full width Sargon A/B search with current vague "ply" (iteration) of top A/B engines and even more vague "ply" of Lc0. Top A/B engines today seem to use their strong eval to guide the search so that it compensates for sparsely searched vague "plies", and after all EBF reductions to come up with vague "plies" as strong as old real plies. Lc0 MCTS (so called MCTS) is even more extreme in this respect, its eval guides the search so well, that it needs a very sparse search to play well.
Its policy (move ordering) guides the search so well.*
Isn't it the same static eval which guides? There are no two nets.
During the TCEC Stockfish-Allie SuFi, Leela seemed equally blind to certain Allie blunders. But Leela would have never played those moves.
Her eval had (still has?) the same blindspots as Allie, but her move ordering vastly superior.
I wonder if SF's move ordering being turned into an 1 ply multiPV Leela search would be a big ELO gainer.
It should theoretically vastly improve SF's move ordering making its search more efficient.
Maybe, interesting, but the search must be guided by the same SF eval. Nobody tried that?
Dann Corbit wrote: ↑Fri May 29, 2020 11:45 pm
I don't think the weak 6 ply searches matter at all.
You can get to 15 plies in an eyeblink.
If you like ultra-hyper-bullet game in one tenth of a second, then use Komodo.
I am not sure how the NN programs would do at ludicrously short time control.
It would be interesting to see an LC0/Komodo match at very high speed.
For reasonable time control on both play and analysis, Stockfish does very well indeed.
Where do you draw the line?
At 12 plies, 15 plies, or even 24 plies?
Don't we want the best possible move(s) at any given time or depth?
If getting the best possible moves at ply 6 impairs getting the best possible moves at ply 30, then I don't want it.
Personally, my minimum analysis depth for a position I am actually interested in is 36 plies for the first pass.
More often, I run 37 ply searches over a file of thousands of positions (it used to be 36 was painful because it took a long time, but the average time for 36 plies is less than a minute now).
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.
