M ANSARI wrote: ↑Mon Apr 29, 2019 5:20 pm
Are you serious ???? Aesthetic ???? Playing on for a 100 moves when a mate in 12 is there is not aesthetics! An engine is an analysis tool that is supposed to help us analyze a chess position so that we can train and learn from it. This really makes Lc0 useless to analyze as a stand alone engine and you will always need another good engine for a sanity check. Time to figure out a way to fix this with Lc0 and if takes removing the "zero" then be it. I just can't believe that we are still at this stage.
Totally agree. I think people lose sight of the fact that the primary use of an engine for most people is as an analysis tool. Engine to engine matches and Elo ratings are a fairly standard way of roughly assessing which engines are best up to the task. But for Lc0 that approach clearly does not work. Lc0 is both spectacular, and spectacularly useless, both at the same time. But it will get there. Where it has got to today is incredible.
Yes of course ... in no way shape or form do I want to denigrate Lc0 or how it has changed computer chess for the better ... it has. I actually went out and bought a new 2080 Ti card just for it. But it is painful to watch these endgames and it almost feels a bit like Rainman ... where the engine is absolutely brilliant in some things ... but can be totally clueless in what we humans consider simple positions ... AI can suffer from autism?? Take for example this position where I will disregard the many earlier moves that are also delaying the win as you can say that earlier moves might prolong the game but are still strong. But now you have this
[d]4Q3/5K2/3k1P2/2b5/8/8/8/8 w - - 0 104
I mean even a 500 ELO player will think ... hmmm ... Queen check on e7 ... take the bishop and promote another queen and mate with queen. Of course a 3500 ELO engine would not do that and will probably find a quicker mate. But Lc0 here took a further 106 moves to mate !!!! I mean come on ... lets just admit that this is not good and try to fix it! Let's not lose sight of the fact that this is about chess engines helping us play better chess rather than just silicone beasts battling things out. Every move that delays mate is an extra move that can be a blunder or cause a crash. If you are on Chess.com they have this new thing called Puzzle Rush and have to admit I am a little addicted to it. If you go for a mate in 3 rather than a mate in 2 ... it buzzes you as wrong answer. Now that is frustrating and I hated that feature ... but recently I have seen that it has helped me make better moves in normal games as well as I will not play the impulsive move and many times there is a stronger move. So yes Lc0 has been amazing and very exciting to see, but God PLEASE someone fix this trolling thing!!!
I can not get my Lc0 to play a trolling endgame. What settings are you using. I tried and it always mates in 19 moves from the starting position...
I am using a 6 man table base. With these settings.
Lc0 Settings.jpg
"The worst thing that can happen to a forum is a running wild attacking moderator(HGM) who is not corrected by the community." - Ed Schröder
But my words like silent raindrops fell. And echoed in the wells of silence.
Even worse than trolling endgames per se, I've seen cases when Lc0 is a few pieces up, it obviously sees that it can win with just its King and a Rook, so it then simply gives up all its other pieces and does exactly that.
M ANSARI wrote: ↑Mon Apr 29, 2019 5:20 pm
An engine is an analysis tool that is supposed to help us analyze a chess position so that we can train and learn from it.
A chess engine is a chess playing program that someone created for some reason. Lc0 is created to see how reinforced learning is doing in chess. Lc0 is for sure NOT created to be an analysis tool for you. If it is not useful for you as analysis tool just use an engine that is better for you, create an own one or wait until someone makes one for you on basis of Lc0. But stop whining. It's getting annoying.
CCRL sources say - new leela that is playing in TCECc up is +37 elo, and multiple testing soureces shows +40 to +50 elo stronger than Lco in last SuFi.
Meanwhile SF gained 0 elo in the last 90 days. https://ccrl.chessdom.com/ccrl/404/cgi/ ... +opponents
Latest result on 1/5/2019 on Spcc, newest leela improved +68 elo from last TCEC Sufi net (T32930).
Leela progress within the last 3 months is pretty amazing! https://www.sp-cc.de/lc0-testing.htm
Nay Lin Tun wrote: ↑Sun May 05, 2019 6:49 am
Latest result on 1/5/2019 on Spcc, newest leela improved +68 elo from last TCEC Sufi net (T32930).
Leela progress within the last 3 months is pretty amazing! https://www.sp-cc.de/lc0-testing.htm
Not only that, but very good gains (60 Elo poitns) since ID41800 from one month ago, where the run seemed to start to drift, and a LR drop seemed necessary. Good that they didn't rush it, and we still hope for even better nets of t40. My prediction for TCEC SuFi: if not very weird openings, Lc0 t40 will crush SF_dev, with some 2:1 W/L ratio. The conditions in TCEC are apparently similar to A0 paper, and I would guess that by now late t40 surpassed significantly the A0 from the paper (on somehow comparable hardware).
M ANSARI wrote: ↑Mon Apr 29, 2019 5:20 pm
An engine is an analysis tool that is supposed to help us analyze a chess position so that we can train and learn from it.
A chess engine is a chess playing program that someone created for some reason. Lc0 is created to see how reinforced learning is doing in chess. Lc0 is for sure NOT created to be an analysis tool for you. If it is not useful for you as analysis tool just use an engine that is better for you, create an own one or wait until someone makes one for you on basis of Lc0. But stop whining. It's getting annoying.
Exactly. The masses are always doing this. They find a new god (car, chess program, pop star, footballer) build him her it up, then start with criticisms and eventually hate as the god is brought down. The masses want gods and heroes (usually out of us) but they also hate them (for being part of them). Probably an endless repeat re-enaction of the child’s disappointment with the father.
Nay Lin Tun wrote: ↑Sun May 05, 2019 6:49 am
Latest result on 1/5/2019 on Spcc, newest leela improved +68 elo from last TCEC Sufi net (T32930).
Leela progress within the last 3 months is pretty amazing! https://www.sp-cc.de/lc0-testing.htm
Not only that, but very good gains (60 Elo poitns) since ID41800 from one month ago, where the run seemed to start to drift, and a LR drop seemed necessary. Good that they didn't rush it, and we still hope for even better nets of t40. My prediction for TCEC SuFi: if not very weird openings, Lc0 t40 will crush SF_dev, with some 2:1 W/L ratio. The conditions in TCEC are apparently similar to A0 paper, and I would guess that by now late t40 surpassed significantly the A0 from the paper (on somehow comparable hardware).
I think that in the last SuFi, Leela was already stronger than SF and by quite a margin, "raw-strength wise". The problem is that lc0 used to still make severe blunders and it threw away a handful of games, be it wins that ended in a draw, or draws that ended in a loss. Apparently, this season is not the same and those quirks have been reduced a lot, or even completely fixed. I don't know how the jhortos nets play in general cause I haven't seen them much, but with official T40 net (if selected) i'm quite sure that Leela will spank SF in the superfinal, at last.
People on this forum have said that LCzero has a 40 MB pre calculated cache of moves. I wish that the TCEC would run an experiment letting SF use the latest Brianfish or else disabling LCzeros cache of moves.
