TalkChess.com

Now it looks like we have the number of a CCRL 2445 engine in 3+2. Only tweak was a little prefetching of high policy children which doubled the nps. I'm tempted to go mutithreaded, but I want to keep the code simple. [Event "?"] [Site "?"] [Date "2020.03.21"] [Round "1"] [White "A0Lite"] [Black "Tu...

Another game against a CCRL ~2150 engine at 3 min + 2 sec. [Event "A0Lite vs Various I: Baislicka - A0Lite"] [Site "https://lichess.org/study/Gj9NmP2G/AlJO1fFS"] [Date "2020.03.20"] [Round "3"] [White "Baislicka"] [Black "A0Lite"] [Result "0-1"] [TimeControl "60+1"] [Termination "adjudication"] [UTC...

I’m putting the finishing touches on my engine, a0lite. It uses the badgyal-8 network — a Frankenstein monster trained up on lichess games and shallow sf multipv search — and is written entirely in python. I’ve been looking for sparring partners that are also written in python. So far the only vague...

I believe this should be the real challenge for programmers now, adding Elo is getting boring (although necessary for some engines). Leela has done something on this front. Small nets and small search result in a very different playing style than a crippled Thinker. https://lichess.org/@/leela2200 ...

All the decisive games available on lichess as studies Nimzo-off 1 thru 7

https://lichess.org/study/o23IJSMk

There were some duplicate games in the nimzo woodpile. Removing those we get: 1) FFv471 41 : 333 (+51,=258,-24), 54.1 % vs. : games ( +, =, -), (%) : Diff, SD, CFS (%) 384x30-3010 : 136 ( 8, 119, 9), 49.6 : +8, 10, 79.1 SF11 : 197 ( 43, 139, 15), 57.1 : +41, 10, 100.0 2) 384x30-3010 33 : 336 (+39,=2...

I’ve put up all the decisive games in the Dorf-off on lichess in 15 studies, Dorf-off 1 thru 15. https://lichess.org/study/UES6Dwmp

The Nimzo-off is complete (sf11cpu8, ff and sv 3010 on rtx 2070). 1) FFv471 37 : 400 (+51,=322,-27), 53.0 % vs. : games ( +, =, -), (%) : Diff, SD, CFS (%) 384x30-3010 : 200 ( 8, 183, 9), 49.8 : +5, 9, 72.6 SF11 : 200 ( 43, 139, 18), 56.2 : +37, 9, 100.0 2) 384x30-3010 32 : 400 (+39,=337,-24), 51.9 ...

Yes, and all the usual games with colors reversed, before anybody asks.

After testing t60 and FF in the Najdorf, I decided to step up to the big nets: Sergio’s 384x30 nets trained on t60 games. I had to extend the TC so the big net got enough nodes. A bit past halfway. Let’s see how we go. 3... Bb4 plus the most popular lines 1 ply after that. TC 3 min + 2 sec. RTX 2070...