Subject: Neural Nets (not Maia) trained on Human Games Fri Jun 09, 2023 7:43 pm
I found some time ago a github feauturing some 256x2 sf nets that will work in much older versions of sf, or engines that use that net architecture. It was
Subject: Re: Neural Nets (not Maia) trained on Human Games Fri Jun 09, 2023 7:44 pm
There are some nets trained on maia, which don't work so well, and a couple king base nets which play excellent human chess, especially when just the net is used, and a net trained on caissabase, which resulted in the strongest net.
Ghppn likes this post
Nezhman
Posts : 74 Join date : 2020-11-27
Subject: Re: Neural Nets (not Maia) trained on Human Games Sat Jun 10, 2023 6:31 am
texium wrote:
I found some time ago a github feauturing some 256x2 sf nets that will work in much older versions of sf, or engines that use that net architecture. It was
What engine could use/run these kennyfrc nets? Thanks!
Ghppn likes this post
texium
Posts : 112 Join date : 2022-07-19
Subject: Re: Neural Nets (not Maia) trained on Human Games Sat Jun 10, 2023 10:48 am
Nezhman wrote:
texium wrote:
I found some time ago a github feauturing some 256x2 sf nets that will work in much older versions of sf, or engines that use that net architecture. It was
What engine could use/run these kennyfrc nets? Thanks!
personally i use Rodent with nnue, as i can create some interesting gm personalities. my nakamura personality, for example, matches his moves online and OTB over 50% of the time. these tuned personalities hover around 50-60% match, which is good because its not too high
Nezhman and Ghppn like this post
texium
Posts : 112 Join date : 2022-07-19
Subject: Re: Neural Nets (not Maia) trained on Human Games Sun Jun 11, 2023 3:37 pm
could someone with a working talkchess account share the github link i shared, to the aforementioned talkchess post
Ghppn likes this post
Nezhman
Posts : 74 Join date : 2020-11-27
Subject: Re: Neural Nets (not Maia) trained on Human Games Sun Jun 11, 2023 11:20 pm
texium wrote:
Nezhman wrote:
texium wrote:
I found some time ago a github feauturing some 256x2 sf nets that will work in much older versions of sf, or engines that use that net architecture. It was
What engine could use/run these kennyfrc nets? Thanks!
personally i use Rodent with nnue, as i can create some interesting gm personalities. my nakamura personality, for example, matches his moves online and OTB over 50% of the time. these tuned personalities hover around 50-60% match, which is good because its not too high
So, the kennyfrc nets do work with Rodent-NNUE, is that confirmed? A net is useless without an engine to run it.
Ghppn likes this post
texium
Posts : 112 Join date : 2022-07-19
Subject: Re: Neural Nets (not Maia) trained on Human Games Mon Jun 12, 2023 1:50 pm
Nezhman wrote:
texium wrote:
Nezhman wrote:
texium wrote:
I found some time ago a github feauturing some 256x2 sf nets that will work in much older versions of sf, or engines that use that net architecture. It was
What engine could use/run these kennyfrc nets? Thanks!
personally i use Rodent with nnue, as i can create some interesting gm personalities. my nakamura personality, for example, matches his moves online and OTB over 50% of the time. these tuned personalities hover around 50-60% match, which is good because its not too high
So, the kennyfrc nets do work with Rodent-NNUE, is that confirmed? A net is useless without an engine to run it.
yes they do! just rename the net you wanna use to the name of the one in the exe folder. also the nets work with any cfish 12 or engine using 20mb sf nets
Nezhman and Ghppn like this post
texium
Posts : 112 Join date : 2022-07-19
Subject: Re: Neural Nets (not Maia) trained on Human Games Mon Jun 12, 2023 2:57 pm
Some things interesting to note: 80 NNUE and 20 Classic, or 65 NNUE and 35 Classic are the best strongest settings, also very human The NNUE actually makes rodent a bit stronger covering many weaknesses in its eval. its at least 3000-3100 but probably the most human-like 3100 you'll ever see. I find rodent with the nn to be more human when i turn off selectivity pruning, making the depth much lower, but a more human search like 90s engines. Here is a game of Rodent using a GM net, playing against Rebel 14 CSTal: [pgn][Event "Test Games"] [Site "Lucas Chess R 2.07a"] [Date "2023.06.12"] [White "Rebel Tal"] [Black "Rodent GM"] [Result "1/2-1/2"] [ECO "B33"] [TimeControl "180+2"] [PlyCount "82"] [Opening "Sicilian: Pelikan (Lasker/Sveshnikov) variation"] [Termination "Draw by Agreement"]
Subject: Re: Neural Nets (not Maia) trained on Human Games Mon Jun 12, 2023 3:26 pm
Here's another great game where the tal personality shreds Nakamura, in tal-like style, also creating a sound, creative novelty in the Snake benoni [pgn][Event "gm games"] [Site "Lucas Chess R 2.06b"] [Date "2023.04.18"] [White "Tal GM"] [Black "Nakamura"] [Result "1-0"] [BlackElo "2800"] [TimeControl "180+5"] [PlyCount "68"] [Termination "Adjudication"]
Subject: Re: Neural Nets (not Maia) trained on Human Games Tue Jun 13, 2023 5:56 am
texium wrote:
Here's another great game where the tal personality shreds Nakamura, in tal-like style, also creating a sound, creative novelty in the Snake benoni [pgn][Event "gm games"] [Site "Lucas Chess R 2.06b"] [Date "2023.04.18"] [White "Tal GM"] [Black "Nakamura"] [Result "1-0"] [BlackElo "2800"] [TimeControl "180+5"] [PlyCount "68"] [Termination "Adjudication"]
Interesting, was 'Nakamura' also a Rodent personality? Curious, where did the book end in the above game?
By the way, I reposted your link to kennyfrc's nets in the talkchess forum!
Ghppn likes this post
texium
Posts : 112 Join date : 2022-07-19
Subject: Re: Neural Nets (not Maia) trained on Human Games Wed Jun 14, 2023 8:12 am
Nezhman wrote:
texium wrote:
Here's another great game where the tal personality shreds Nakamura, in tal-like style, also creating a sound, creative novelty in the Snake benoni [pgn][Event "gm games"] [Site "Lucas Chess R 2.06b"] [Date "2023.04.18"] [White "Tal GM"] [Black "Nakamura"] [Result "1-0"] [BlackElo "2800"] [TimeControl "180+5"] [PlyCount "68"] [Termination "Adjudication"]
Interesting, was 'Nakamura' also a Rodent personality? Curious, where did the book end in the above game?
By the way, I reposted your link to kennyfrc's nets in the talkchess forum!
no its not. its one i made based on the fischer personality, then tuned through move matching of hikaru's OTB and online games. Over 300 moves at d8, it gets at least 50% which isnt bad
Nezhman likes this post
texium
Posts : 112 Join date : 2022-07-19
Subject: Re: Neural Nets (not Maia) trained on Human Games Wed Jun 14, 2023 8:25 am
Nezhman wrote:
texium wrote:
Here's another great game where the tal personality shreds Nakamura, in tal-like style, also creating a sound, creative novelty in the Snake benoni [pgn][Event "gm games"] [Site "Lucas Chess R 2.06b"] [Date "2023.04.18"] [White "Tal GM"] [Black "Nakamura"] [Result "1-0"] [BlackElo "2800"] [TimeControl "180+5"] [PlyCount "68"] [Termination "Adjudication"]
Interesting, was 'Nakamura' also a Rodent personality? Curious, where did the book end in the above game?
By the way, I reposted your link to kennyfrc's nets in the talkchess forum!
Book ended by move 4 which makes the 6. g4 novelty a really interesting move thats essentially unplayed entirely besides a few lichess games.
Nezhman likes this post
Nezhman
Posts : 74 Join date : 2020-11-27
Subject: Re: Neural Nets (not Maia) trained on Human Games Wed Jun 14, 2023 11:41 pm
texium wrote:
Nezhman wrote:
texium wrote:
Here's another great game where the tal personality shreds Nakamura, in tal-like style, also creating a sound, creative novelty in the Snake benoni [pgn][Event "gm games"] [Site "Lucas Chess R 2.06b"] [Date "2023.04.18"] [White "Tal GM"] [Black "Nakamura"] [Result "1-0"] [BlackElo "2800"] [TimeControl "180+5"] [PlyCount "68"] [Termination "Adjudication"]
Home 