search

lightvector / leela-analysis

Analysis scripts for Leela
Other
56 stars 13 forks source link

Quick win: horizon effect #10

Open ThomasCabaret opened 6 years ago

ThomasCabaret commented 6 years ago

Hi, there is a clear horizon effect on hard games, by example: black win rate = 50% white makes a move. then black win rate drops to 20%. It comes from the fact the analysis is probably done forward but performing the analysis backward would fix this kind of issue (because leela engine keeps track of explored sub-tree), it will not make leela that much reliable but it will improve it's evaluation with the game main line as an hint. And more importantly it makes the analysis more consistent with itself. So performing the evalution backward from last move to first move is probably a real quick win for the script!

lightvector commented 6 years ago

Cool idea. Could you PM me with an example SGF file with a horizon effect so that I can test it out?

On Tue, Jan 23, 2018 at 9:06 AM, ThomasCabaret notifications@github.com wrote:

Hi, there is a clear horizon effect on hard games, by example: black win rate = 50% white makes a move. then black win rate drops to 20%. It comes from the fact the analysis is probably done forward but performing the analysis backward would fix this kind of issue (because leela engine keeps track of explored sub-tree), it will not make leela that much reliable but it will improve it's evaluation with the game main line as an hint. And more importantly it makes the analysis more consistent with itself. So performing the evalution backward from last move to first move is probably a real quick win for the script!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/lightvector/leela-analysis/issues/10, or mute the thread https://github.com/notifications/unsubscribe-auth/ALY5-wsqm1i6aTnUduMC_4MqP7C4_8twks5tNeeAgaJpZM4Rpspe .

ThomasCabaret commented 6 years ago

You can download an example here: http://eidogo.com/#4cYWS2Fgy move 119-120 a drop of 30% for black during white's move. I tested leela empty from all evaluation starting on move 120 then going back to move 119. It uses the subtree explored in 120 (evaluates 119 clearly faster when it has seen 120 first) but it's not that instantaneous (I dont know why, I ll see with the leela dev) It seems to need something like 10 seconds. But proceeding backward you can force the consistency by waiting for evalution move N to reach something "close or better" to the one move N+1.