>what kind of large pawn patterns did You manage to implement?
main large pawn pattern was this:
all possible combinations of 2 pawns of the same color.
E.G., white pawn on A3 and another white pawn on B3 is feature #1, with a weight;
white pawn on A4 and another white pawn on B3 is feature #2, with a different weight; continue for all possible squares of both pawns.
I think this helped Fizbo improve endgame (handling of multiple passed pawns), partly because than endgame was pretty weak to begin with.
Fizbo 1.7 is released
Moderators: hgm, Rebel, chrisw
-
- Posts: 64
- Joined: Fri Oct 18, 2013 11:40 pm
- Location: New York
-
- Posts: 2283
- Joined: Sat Jun 02, 2012 2:13 am
Re: Fizbo 1.7 is released
That's awesome, Youri - thanks!!
Cheers,
CL
Cheers,
CL
ymatioun wrote:Fizbo version 1.7 has just been released and can be downloaded from https://sites.google.com/site/fizbochessengine/.
Changes from version 1.6 are:
1. Add support for 6-piece SYZYGY tablebases, including option "EGTBProbeLimit" - set to 5 by default - that allows one to only use 5-pieces (or 4-piece) tables, even if 6-piece tables are present. I tried adding 6-piece logic before, but could never get it to work right. But now i used working implementation from Stockfish 7 as an example, and made 6-piece syzygy work for me.
2. Extensive changes to evaluation function: change existing evaluation features, add multiple new features (including large pawn patterns), change the way evaluation coefficients are tuned. Estimated increase in playing strength is >50 ELO. Now i estimated playing strength by playing gauntlet against several opponents, so increase in playing strength is real (in the past i mostly played against earlier version of Fizbo, and that often overstated increase in playing strength).
3. search and parallel search logic is mostly unchanged.
Youri.
-
- Posts: 893
- Joined: Mon Jan 15, 2007 11:23 am
- Location: Warsza
Re: Fizbo 1.7 is released
Sounds interesting. Rodent 1.7 had some similar bishop + pawn patterns. I believe there's a lot to gain from this kind of knowledge.
Pawel Koziol
http://www.pkoziol.cal24.pl/rodent/rodent.htm
http://www.pkoziol.cal24.pl/rodent/rodent.htm
-
- Posts: 1539
- Joined: Thu Mar 09, 2006 2:02 pm
Re: Fizbo 1.7 is released
Thanks for a new version but ... (Thats a bad start for a sentence )
Watching Fizbo play chess is hard.
1. PV is only 2 plys deep
2. No Nodes per second
3. No current move (which means that on longer games the engine apears dead in a GUI as it needs longer from ply 25 to 26)
and a bug: It doesn't ponder on ponderhits. (If it gets an immediate ponderhit it answers immediately and waits for the opponents move, same as early Anscacs versions). Thats worth a few elo! Easy to check, start a medium length "ponder on" game (10m) and watch the CPU load in the task manager ... sometimes while Fizbo could ponder it sits idle at 0 percent.
That is in order of "annoyance"
Thanks again
Ingo
Watching Fizbo play chess is hard.
1. PV is only 2 plys deep
2. No Nodes per second
3. No current move (which means that on longer games the engine apears dead in a GUI as it needs longer from ply 25 to 26)
and a bug: It doesn't ponder on ponderhits. (If it gets an immediate ponderhit it answers immediately and waits for the opponents move, same as early Anscacs versions). Thats worth a few elo! Easy to check, start a medium length "ponder on" game (10m) and watch the CPU load in the task manager ... sometimes while Fizbo could ponder it sits idle at 0 percent.
That is in order of "annoyance"
Thanks again
Ingo
-
- Posts: 64
- Joined: Fri Oct 18, 2013 11:40 pm
- Location: New York
Re: Fizbo 1.7 is released
I'll look into these items and see if i can address them in the next version. Off the top of my head now:
1. short PV: this is a known limitation. I never felt a need for longer PV, so i never bothered to make it longer.
2. in Winboard, GUI shows NPS, calculated from nodes and seconds, so i never missed NPS. But this is easy to add, so i'll add it.
3. currently i only print output when new best move is found or new depth is reached. I'l look into adding more frequent output.
Pondering on ponderhits: i think i never implemented this logic; again, i'll look into this.
Youri.
1. short PV: this is a known limitation. I never felt a need for longer PV, so i never bothered to make it longer.
2. in Winboard, GUI shows NPS, calculated from nodes and seconds, so i never missed NPS. But this is easy to add, so i'll add it.
3. currently i only print output when new best move is found or new depth is reached. I'l look into adding more frequent output.
Pondering on ponderhits: i think i never implemented this logic; again, i'll look into this.
Youri.
-
- Posts: 1539
- Joined: Thu Mar 09, 2006 2:02 pm
Re: Fizbo 1.7 is released
Hello
As your engine seems to enter the top 15 range the points 1 and 3 are needed for everyone who wants to use your engine for analysis. Without that, the engine can be very boring to check a position.
Point 2 however is different. Shredder Classic GUI doesnt calculate because there is an "info nps <x>" in UCI for this (defined as engine "should send"!). Especially in combination with a more frequent point 3 this makes very much sense.
I am looking forward to the next Fizbo then
Ingo
PS: hashfull <x> and seldepth <x> are nice too if the engfine is used for analysis
Just two remarks:ymatioun wrote: 1. short PV: this is a known limitation. I never felt a need for longer PV, so i never bothered to make it longer.
2. in Winboard, GUI shows NPS, calculated from nodes and seconds, so i never missed NPS. But this is easy to add, so i'll add it.
3. currently i only print output when new best move is found or new depth is reached. I'l look into adding more frequent output.
Pondering on ponderhits: i think i never implemented this logic; again, i'll look into this.
As your engine seems to enter the top 15 range the points 1 and 3 are needed for everyone who wants to use your engine for analysis. Without that, the engine can be very boring to check a position.
Point 2 however is different. Shredder Classic GUI doesnt calculate because there is an "info nps <x>" in UCI for this (defined as engine "should send"!). Especially in combination with a more frequent point 3 this makes very much sense.
I am looking forward to the next Fizbo then
Ingo
PS: hashfull <x> and seldepth <x> are nice too if the engfine is used for analysis
-
- Posts: 64
- Joined: Fri Oct 18, 2013 11:40 pm
- Location: New York
Re: Fizbo 1.7 is released
I understand, and will implement as much of this as i can for the next version.
Thanks for suggestions.
Youri.
Thanks for suggestions.
Youri.
-
- Posts: 41428
- Joined: Sun Feb 26, 2006 10:52 am
- Location: Auckland, NZ
Re: Fizbo 1.7 is released
From what I've seen so far, +50 Elo seems conservative.ymatioun wrote:Fizbo version 1.7 has just been released and can be downloaded from https://sites.google.com/site/fizbochessengine/.
.....Estimated increase in playing strength is >50 ELO.
Youri.
gbanksnz at gmail.com
-
- Posts: 64
- Joined: Fri Oct 18, 2013 11:40 pm
- Location: New York
Re: Fizbo 1.7 is released
I hope it is more. But in the past i overstated strength improvements several times, so i wanted to avoid it now.
I did my testing at shorter time controls (40/4, running 8 games at a time on 4 cores with hyper-threading), and it showed around +100 ELO vs Fizbo 1.6. But i was not sure this would hold up at longer time controls.
Youri.
I did my testing at shorter time controls (40/4, running 8 games at a time on 4 cores with hyper-threading), and it showed around +100 ELO vs Fizbo 1.6. But i was not sure this would hold up at longer time controls.
Youri.