Question for Norman Schmidt ... Fizbo is the topic!

[Frank Quisinsky]

Hi Norman,

you have created Hakkapeliitta NN
https://github.com/FireFather?tab=repositories

You know ... Fizbo is here:
https://github.com/MoonstoneLight/Fizbo-Chess



Is it possible for you to do the same for Fizbo?

At the moment I am working on the conditions for a new tournament with 41 engines.
Can be see in the Excel from my still running FCP-Tourney-2024.
https://www.amateurschach.de/download/_ ... y-2024.zip

Different from the older HCE engine will be inside.
All engines with a too high move average / move average wins from the still-running tournament will be replaced.
Your Hakkapeliitta is in the group of 41 engines, Fizbo 2 too... but to have an NN version would be great.

Changes in my conditions:

1. Time control = 60 moves in 60 minutes + 6
2. Time control = all moves in 30 minutes + 3
6-pieces syzygy-bases
Hashtables = 1Gb
Again without resing-mode!

The time control is the reason I am looking for engines that produce a "normal" move average.

Again, Fizbo 2 NN would be great!!

Best
Frank

[Frank Quisinsky]

Hi Norman,

ah, I forgot ...
I will start the new tournament before Xmas 3023!
If not possible ... Fizbo 2 can play without NN, only a question!

Best
Frank

[carldaman]

I thought Fizbo2 was already using a NN.

[connor_mcmonigle]

Fizbo was among the first strong chess engines to use a neural network for position evaluation (several years ahead of its time). The network is very small and quite weak by modern standards (the plain text weights are hardcoded in a source file: https://github.com/MoonstoneLight/Fizbo ... nn.cpp#L86). It would really be a shame if Fizbo's unique evaluation function was ripped out and replaced with a network and some code copied from Stockfish.

However, it would be neat if the code was formatted + modernized and made portable (swapping out YBWC for LazySMP would also be a nice contribution imo).

[kranium]

Fizbo was among the first strong chess engines to use a neural network for position evaluation (several years ahead of its time). The network is very small and quite weak by modern standards (the plain text weights are hardcoded in a source file: https://github.com/MoonstoneLight/Fizbo ... nn.cpp#L86). It would really be a shame if Fizbo's unique evaluation function was ripped out and replaced with a network and some code copied from Stockfish.

However, it would be neat if the code was formatted + modernized and made portable (swapping out YBWC for LazySMP would also be a nice contribution imo).

Hakepelitta's NNUE implementation uses Daniel Shawul's nnue-probe code
https://github.com/dshawul/nnue-probe/
It was not copied from Stockfish, but was created by Ronald de Man.
"The core nnue probing code is taken from CFish and modified a bit."

The intention of Ronald and Daniel's releases were to simplify implementation and to share and facilitate the use of NNUE in other engines.
Clearly a ~200 Elo Fizbo improvement could be enjoyed by many hobbyists.

I measured move selection of Fire 8.NN (using Raptor, the strong SV NNUE from SF12) against the last 4 versions of Stockfish...result:
Very unique play...Fire made the same move as SF only 43.34% of the time.
https://github.com/FireFather/fire (see bottom of page)

A Fizbo NNUE implementation using nnue-probe wouldn't damage Fizbo in any way...it would just create a different & much stronger flavor.
Why this would be a 'shame' is beyond me.? (other than the fact that it might be as strong or better than Seer )
I guess that would depend on which net you choose...there are many choices, here are a few:
https://github.com/FireFather/halfkp_256x2-32-32-nets

Anyway, if it bothers you, you could just not download Fizbo NNUE...and simply choose to utilize Fizbo in it's original form.

[kranium]

Anyway, I already answered Frank privately several hours ago...

"Thanks for the suggestion, but I'm not too interested in Fizbo
it's a really really big codebase that needs a ton of work
the coding is rough, haphazard and confusing IMHO"

Perhaps someone else will take up the challenge...

[connor_mcmonigle]

...
A Fizbo NNUE implementation using nnue-probe wouldn't damage Fizbo in any way...it would just create a different & much stronger flavor.
...

Because you'd be removing one of the most interesting and original aspects of Fizbo both from a playing style perspective and an engineering perspective. You'd be lopping off its pioneering evaluation function and replacing it with Stockfish ~12's evaluation function. Whether the code is copied from Stockfish or copied from the reimplementation of the Stockfish code in CFish could not possibly be less relevant. Fizbo was renowned for its unique playing style and if you're going to argue that its evaluation function had no bearing on its playing style, you're delusional.

Separately, I'd be surprised if it proved stronger than Seer at this point, but I'm also not really inclined to care.

Edit: To clarify, the existence of a Fizbo+SF-NNUE wouldn't be a shame. The shame would be in anyone (such as Frank) actually using it.

[Frank Quisinsky]

Hi Norman, Connor,

thanks for the information in addition I found in your comments.
Good known that for me is the playing style of an chess engine is most important and have the first priority.

Such records are for me a K.O. criteria for each engine ...
Thanks to be god ... only a small group of engines!

Move-average draw games:

Stockfish 16 NN
white draw 110 moves
black draw 114 moves ... a new record!

Koivisto 9.2 NN dev
white draw 120 moves
black draw 102 moves

Booot 7.2
white draw 113 moves
black draw 110 moves

Berserk 12
white draw 128 moves ... a new record!
black draw 106 moves

Only a small group of programs do the same!
let us compare with ...

Seer 2.6.0 NN
white draw 86
black draw 92

Seer 2.7.0 NN
white draw 92
black draw 86

Looks normal ...
Or with the the number 5 or 6

CSTal 2.0 NN
white draw 85
black draw 83

Or with

Dragon by Komodo 3.3 (Contempt = 0):
white draw 93
black draw 92

So I am looking for nice engines can replace the the group of "Playing Style Killers" ...
You know, more nice for looking while the games are still running.

What Booot, Berserk, Stockfish, Kallisto and different others do here ...
No chess player like to see that while a draw game is still-running.
That's the same as Kg1-h1 in beginning of the first chess-computers.
For me an absolutely K.O. criteria for watching the games.

I am thinking at first ... maybe Fizbo can be make a bit stronger.
The produced playing style is just great!

Best
Frank

[Frank Quisinsky]

Hi Norman,

yes, Hakkapeliitta is great with the NN, a nice work you do here!
Will test it a bit ... interesting are the differents to Andscacs 0.1 NN dev with the same NN file (both can use the same NN file).

Move average from Andscacs 0.1 NN is just great (for wins and draws).
And not such a cramp Stockfish and Co. do for draw games.

Best
Frank

[Frank Quisinsky]

Hi there,

from Excel ...



Must replace a bit with longer time-controls I like to use for the next tourney ... or I lose to many time for nothing.

Excel can be found in the download file from my still-running tourney.

Best
Frank