| Life In 19x19 http://www.lifein19x19.com/ |
|
| LZ help for position http://www.lifein19x19.com/viewtopic.php?f=18&t=16162 |
Page 1 of 5 |
| Author: | Pippen [ Fri Oct 19, 2018 6:48 am ] |
| Post subject: | LZ help for position |
Can somebody run LZ for an analysis of the position below (Black to play) to check the difference with Leela 11 (see pic)? Just a couple of minutes. I hope LZ also offers something similiar of an Analysis Window like Leela to see deeper into its thought process. Attachment: Leela.jpg [ 252.05 KiB | Viewed 17227 times ] |
|
| Author: | Bill Spight [ Fri Oct 19, 2018 7:18 am ] |
| Post subject: | Re: LZ help for position |
Looks like it's hard for Black to go wrong in this position. 
|
|
| Author: | Pippen [ Fri Oct 19, 2018 7:51 am ] |
| Post subject: | Re: LZ help for position |
Bill Spight wrote: Looks like it's hard for Black to go wrong in this position. Well, first it's interesting how LZ evaluates this position since Leela thinks it's White's favour already. Secondly, I like to see what LZ proposes: g3/f3 or d5/c5 or something completely different. I think d5 is best here, but I have a feeling a brutal tenuki might be even better. |
|
| Author: | Tryss [ Fri Oct 19, 2018 7:55 am ] |
| Post subject: | Re: LZ help for position |
A LZ analysis at around 100k playouts with network 181 : Note that on an empty board, Black has 46.5% winrate (basically, it didn't move much, the splitting move was a little over -2%, but the white answer balanced this |
|
| Author: | Pippen [ Fri Oct 19, 2018 8:47 am ] |
| Post subject: | Re: LZ help for position |
Thx a lot. Wow, would never have thought of playing on the right lower side.  Is 100K playouts much or little in terms of analysis depth?
|
|
| Author: | Tryss [ Fri Oct 19, 2018 9:07 am ] |
| Post subject: | Re: LZ help for position |
It's a lot. LZ training games and matches are played at 3200 playouts. |
|
| Author: | Uberdude [ Fri Oct 19, 2018 12:03 pm ] |
| Post subject: | Re: LZ help for position |
Pippen: 1 LZ playout is worth a lot more than 1 Leela simulation (and not just because LZ is stronger), they aren't equivalent as a "unit of work". Maybe a factor of 10 difference? Also LZ's principal variations stay sensible far deeper into the tree whereas Leela's tend to turn to junk maybe 6-10 moves deep. |
|
| Author: | Pippen [ Fri Oct 19, 2018 1:28 pm ] |
| Post subject: | Re: LZ help for position |
Tryss wrote: Note that on an empty board, Black has 46.5% winrate (basically, it didn't move much, the splitting move was a little over -2%, but the white answer balanced this Maybe interesting to see what White should have answered instead of M3 and Leela's ideas. In any way my Split-Fuseki seems far more competitive than even I thought. It basically loses nothing, even in a scenerio like here where virtually nobody except me ever plays it (one pro game ever^^). I can't wait till LZ is available without all that installing mambo-jambo to explore it myself. BTW: Any idea when LZ might be available like Leela 11 with or without payment? Did the developer say something? |
|
| Author: | Tryss [ Fri Oct 19, 2018 3:05 pm ] |
| Post subject: | Re: LZ help for position |
LZ split fuseki possible continuation : But what Installing mambo-jumbo are you talking about? If you use windows and have a GPU Step 1) download Lizzie (the windows GPU version) and unzip it to the folder you want : https://github.com/featurecat/lizzie/releases/tag/0.5 Launch lizzie.jar to start LZ with a nice interface Step 2 (optionnal, but won't hurt) download and replace the Network file with the latest one. You can get it here : http://zero.sjeng.org/best-network |
|
| Author: | Gomoto [ Fri Oct 19, 2018 3:25 pm ] |
| Post subject: | Re: LZ help for position |
Attachment: badforblack.jpg [ 154.89 KiB | Viewed 17112 times ] |
|
| Author: | Kirby [ Fri Oct 19, 2018 6:32 pm ] |
| Post subject: | Re: LZ help for position |
Elf v1 prefers black here (prefers black from empty board, too), but recommends the same move as LZ: |
|
| Author: | Kirby [ Fri Oct 19, 2018 6:37 pm ] |
| Post subject: | Re: LZ help for position |
Also, even though Elf likes black at the beginning, it thinks black's split is a bad move, and then thinks white is winning: |
|
| Author: | Kirby [ Fri Oct 19, 2018 6:41 pm ] |
| Post subject: | Re: LZ help for position |
All this being said, I will say that I've been screwing around with Elf lately, and I noticed when replaying the Lee Sedol - AlphaGo game 4, where Lee Sedol won... Elf obviously thinks AlphaGo is majorly winning around the middle of the game, but after Lee Sedol's "divine move" (of 78), which doesn't actually work, Elf still thinks AlphaGo is winning. And while it's not Elf's top choice, AlphaGo's response still has a 95% or so winrate. In other words, even after showing Elf AlphaGo's mistake, Elf, like AlphaGo, still thinks that black was winning the game. As the game plays out, like AlphaGo, Elf's winrate dropped significantly when it finally realized that the result had changed. So these types of analyses are interesting, but it's also interesting to note the mistakes that they make sometimes. |
|
| Author: | Pippen [ Fri Oct 19, 2018 11:47 pm ] |
| Post subject: | Re: LZ help for position |
Tryss wrote: What evaluation does LZ give Black after 6 and then for the position in the diagram? (remark for me: 46.5% for Black on an empty board) |
|
| Author: | Pippen [ Fri Oct 19, 2018 11:52 pm ] |
| Post subject: | Re: LZ help for position |
Tryss wrote: Step 1) download Lizzie (the windows GPU version) and unzip it to the folder you want : https://github.com/featurecat/lizzie/releases/tag/0.5 Launch lizzie.jar to start LZ with a nice interface I have Windows Vista and as far as I know it's not working there. So I would have to go some way to make LZ working...too much effort. |
|
| Author: | Kirby [ Sat Oct 20, 2018 4:28 am ] |
| Post subject: | Re: LZ help for position |
Lizzie makes it easy https://github.com/featurecat/lizzie/releases |
|
| Author: | Pippen [ Tue Oct 23, 2018 4:34 pm ] |
| Post subject: | Re: LZ help for position |
Black to play. How does LZ see Black and what's LZ favourite continuations? I played at 'a' by the way. |
|
| Author: | Tryss [ Tue Oct 23, 2018 5:30 pm ] |
| Post subject: | Re: LZ help for position |
Game is still equal (black winrate is at 46.9%) LZ wants to attach at C4, and her second choices are to make a shimari in the top left (both directions are considered C13, C14, F17 or G17). Your move is not considered at all. When played, it's a -3% winrate drop, and LZ expects white to tenuki. My guess is that the right side is not very important right now (it's kinda flat?) LZ first main line After that, white has several options at a, b or c. In the original sequence (when evaluating the board before 1) she choose b, but after over 15k playouts before playing 4, the first choice switched to a. Final position (with possible white following as x) : The board is still equal (47% for black) |
|
| Author: | Pippen [ Tue Oct 30, 2018 7:53 pm ] |
| Post subject: | Re: LZ help for position |
I am interested in the sequenze 8-32 purely for White. 1. Where is the first time Leela would have played differently as White. 2. After 32, if Black plays F8 White cannot kill the 6 black stones inside IMO. What sequenze would Leela play to F8 and how does it see White then? |
|
| Author: | Waylon [ Wed Oct 31, 2018 1:38 pm ] |
| Post subject: | Re: LZ help for position |
Pippen wrote: I am interested in the sequenze 8-32 purely for White. 1. Where is the first time Leela would have played differently as White. 2. After 32, if Black plays F8 White cannot kill the 6 black stones inside IMO. What sequenze would Leela play to F8 and how does it see White then? Starting with move 8, the picture shows which moves Leela Zero would play differently for White. (net #184) The main line suggested by Leela can be seen in the small board on the left. |
|
| Page 1 of 5 | All times are UTC - 8 hours [ DST ] |
| Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group http://www.phpbb.com/ |
|
, white play at a or b
