Funny thing, I considered Growing_Fruit the most useless piece of chess code I have written. Nice that it found some use.
I might do some fiddling with the search - not much, because recently I am trying to understand the inner workings of go engines. It's a strange field, MCTS search seems very competent strategically but making it understand simple tactics is a pain.
Also, what would You need to create a network based on Rodent Tal? Would games in pgn format be enough? How many?
Admin and matejst like this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: Rebel 14 Tue Jan 18, 2022 1:43 pm
It would be great if you cooperate for OUR good (yes, I know, it is selfish) and kindly offer us a faster search, multiPV for Rebel, and a RodentTal net.
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: Rebel 14 Tue Jan 18, 2022 1:47 pm
Admin wrote:
Interesting sacrifice but reckless
It happens very often in games. In analysis, with a bit more time, Rebel switches to safer continuations.
Then, from my experience, the new net plays endings very well, managing to draw inferior positions, swiftly converting won positions. Very good with handling of free pawns.
nescitus
Posts : 39 Join date : 2020-12-01
Subject: Re: Rebel 14 Tue Jan 18, 2022 2:38 pm
First meaningful improvement: use 75% of NNUE eval value instead of 100%, scores roughly 57% vs original. This value sits better with pruning/reduction margins. Percentage should be exposed as an option, as further tuning is probably possible. Output from LittleBlitzer:
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: Rebel 14 Tue Jan 18, 2022 4:06 pm
nescitus wrote:
Funny thing, I considered Growing_Fruit the most useless piece of chess code I have written. Nice that it found some use.
I might do some fiddling with the search - not much, because recently I am trying to understand the inner workings of go engines. It's a strange field, MCTS search seems very competent strategically but making it understand simple tactics is a pain.
Also, what would You need to create a network based on Rodent Tal? Would games in pgn format be enough? How many?
Do you happen to remember, or know, if popcount instruction gets used anywhere in GrowingFruit?
Admin Admin
Posts : 2528 Join date : 2020-11-17 Location : Netherlands
After some analysis with SF the move isn't losing nor winning, just another case of lack of depth.
Admin Admin
Posts : 2528 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 14 Tue Jan 18, 2022 5:24 pm
nescitus wrote:
First meaningful improvement: use 75% of NNUE eval value instead of 100%, scores roughly 57% vs original. This value sits better with pruning/reduction margins. Percentage should be exposed as an option, as further tuning is probably possible. Output from LittleBlitzer:
Growing Fruit is written in C++ (link). Isn't it possible to just search the source code for "popcount"?
Also, in a high level language, one would expect a method to be implemented in software, but using hardware where available.
I think I might be missing something here?
nescitus
Posts : 39 Join date : 2020-12-01
Subject: Re: Rebel 14 Tue Jan 18, 2022 5:30 pm
Chris Whittington wrote:
Do you happen to remember, or know, if popcount instruction gets used anywhere in GrowingFruit?
All popcounts are found in nnue.h. Another problem with compiling might be caused by all these flavours of intrin.h
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: Rebel 14 Tue Jan 18, 2022 5:45 pm
TheSelfImprover wrote:
Growing Fruit is written in C++ (link). Isn't it possible to just search the source code for "popcount"?
Also, in a high level language, one would expect a method to be implemented in software, but using hardware where available.
I think I might be missing something here?
It might have been named countbits or just about anything. The intrinsic used is subject to all manner of weird possible namings. One doesn’t know what to look for, sure one can eventually find something (or not) but it’s not a high priority and not worth spending hunting time on.
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: Rebel 14 Tue Jan 18, 2022 5:47 pm
nescitus wrote:
Chris Whittington wrote:
Do you happen to remember, or know, if popcount instruction gets used anywhere in GrowingFruit?
All popcounts are found in nnue.h. Another problem with compiling might be caused by all these flavours of intrin.h
Haha! That’s probably me, Linux and Microsoft and all those various systems it had been compiled for in the past.
Admin Admin
Posts : 2528 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 14 Tue Jan 18, 2022 6:30 pm
nescitus wrote:
Funny thing, I considered Growing_Fruit the most useless piece of chess code I have written. Nice that it found some use.
Exactly what I needed, if only not to be accused of stealing search elo
Quote :
I might do some fiddling with the search - not much, because recently I am trying to understand the inner workings of go engines. It's a strange field, MCTS search seems very competent strategically but making it understand simple tactics is a pain.
Also, what would You need to create a network based on Rodent Tal? Would games in pgn format be enough? How many?
Sounds interesting, I prefer EPD, at least 500 million of them and more is welcome, with score and game result, in any EPD format, and with low depth, say depth=7, 8 or 9 as maximum.
nescitus
Posts : 39 Join date : 2020-12-01
Subject: Re: Rebel 14 Tue Jan 18, 2022 6:45 pm
Admin wrote:
Sounds interesting, I prefer EPD, at least 500 million of them and more is welcome, with score and game result, in any EPD format, and with low depth, say depth=7, 8 or 9 as maximum.
Give me one line example how these data should look.
Also, one more working patch can be applied at search.full, line 496 onwards:
Posts : 2528 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 14 Tue Jan 18, 2022 7:16 pm
nescitus wrote:
Admin wrote:
Sounds interesting, I prefer EPD, at least 500 million of them and more is welcome, with score and game result, in any EPD format, and with low depth, say depth=7, 8 or 9 as maximum.
Give me one line example how these data should look.
Code:
rnbqkbnr/p1p2ppp/4p3/1B1p4/1P6/4P2P/P1PP1PP1/RNBQK1NR b KQkq - 0 2 -7 c7c6 -1 rnbqkbnr/p4ppp/2p1p3/1B1p4/1P6/4P2P/P1PP1PP1/RNBQK1NR w KQkq - 0 3 14 b5e2 +1 rnbqk1nr/p4ppp/2p1p3/3p4/1b6/4P2P/P1PPBPP1/RNBQK1NR w KQkq - 0 5 11 g1f3 +1
rnbqkbnr/p1p2ppp/4p3/1B1p4/1P6/4P2P/P1PP1PP1/RNBQK1NR b KQkq - 0 2 -7 c7c6 -1
0 = counter for fifty move rule, I don't use it, always 0. 2 = ply number in game, I don't use it either. -7 = score of move -0.07 in this case. c7c6 = move in long notation. -1 = game result site to move. So in this case black loses (-1) but next move (b5e2) wins (+1), draws are (0).
One step further which greatly reduces the size: 1. Skip captures, including EP; 2. Skip promotions; 3. Skip king in check situations; 4 Skip moves that give check;
As you can see from the 3 example epds there is no ply=4 epd, likely it was a capture.
That would be helpful.
Quote :
Also, one more working patch can be applied at search.full, line 496 onwards:
Posts : 2528 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 14 Tue Jan 18, 2022 7:30 pm
BTW Pawel, what happened to your Rodent NNUE version, I don't see it tested on CEGT and CCRL.
nescitus
Posts : 39 Join date : 2020-12-01
Subject: Re: Rebel 14 Tue Jan 18, 2022 7:39 pm
Admin wrote:
BTW Pawel, what happened to your Rodent NNUE version, I don't see it tested on CEGT and CCRL.
Nothing. Release is unofficial and strength gain comes from Stockfish. No need to test it, doesn't matter if it is 3200 or 3300 Elo.
Admin Admin
Posts : 2528 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel 14 Wed Jan 19, 2022 10:56 am
Pawel, one other search change I am intended to add to growing fruit is to replace IID with IIR which gave ProDeo an extra ply and ~35 elo. It's described here. It even gave elo in Toga, maybe in yours as well. Many engines use it in the mean time in various ways.
Home 
