I really do NOT enjoy playing Leela as it lacks a killer instinct to shorten the game with least moves to mate, etc.
When I slip with my program, it absolutely crushes me quickly.
I find that style of play MUCH more attractive and chess-like than something like a Petrosian on steroids.
Disappointing!
If anyone knows of a Leela network which does NOT have that problem and has the FIGHTING EDGE in tactical melees rather than
always postponing gains for greater gains, but instead CONVERTING gains to substantive crushes, do let me know.
Cheerio.
playing Talish (my program) as opposed to playing Leela
Moderators: hgm, Rebel, chrisw
-
- Posts: 737
- Joined: Wed Mar 08, 2006 8:08 pm
- Location: Orange County California
- Full name: Stuart Cracraft
-
- Posts: 536
- Joined: Thu Mar 09, 2006 3:01 pm
Re: playing Talish (my program) as opposed to playing Leela
This has been a recurring topic on the Leela chess Discord for a long time.
The nets have been trained to win with no extra preference given to a shorter win.
In the past this was characterized as Leela "trolling".
Recently, a lot more attention has been focused on this area (MLH and other approaches), although a solution is still being developed; then, new nets will have to be trained with that additional goal.
The Leela chess project continues to rapidly advance, but this issue will take some time to resolve.
Time frame looks like at least weeks if not months (for a very strong net).
In the mean time, hybrid approaches are getting another look.
Lc0 for opening and midgame, and SF (or other) A/B engines with tablebases for near optimal play in endgames.
Of course, where the hand-off actually happens is not trivial to determine.
The nets have been trained to win with no extra preference given to a shorter win.
In the past this was characterized as Leela "trolling".
Recently, a lot more attention has been focused on this area (MLH and other approaches), although a solution is still being developed; then, new nets will have to be trained with that additional goal.
The Leela chess project continues to rapidly advance, but this issue will take some time to resolve.
Time frame looks like at least weeks if not months (for a very strong net).
In the mean time, hybrid approaches are getting another look.
Lc0 for opening and midgame, and SF (or other) A/B engines with tablebases for near optimal play in endgames.
Of course, where the hand-off actually happens is not trivial to determine.
-
- Posts: 97
- Joined: Mon Jun 25, 2012 10:16 pm
- Location: Forks, WA
- Full name: Ben Nye
Re: playing Talish (my program) as opposed to playing Leela
How about this, run a search with stockfish in parallel, using half the available cores(so that the leela search won't be hurt much at all, it should just need a few cores plus the gpu), and when the stockfish eval passes +8 it's time to hand it over?
-
- Posts: 1242
- Joined: Sat Jul 05, 2014 7:54 am
- Location: Southwest USA
Re:Talish Chess Engine?
Talish Chess Engine/Program?....No search results on a Chess Engine by the name of "Talish" but there is "OpenTal"smcracraft wrote: ↑Fri Apr 17, 2020 8:14 am I really do NOT enjoy playing Leela as it lacks a killer instinct to shorten the game with least moves to mate, etc.
When I slip with my program, it absolutely crushes me quickly.
I find that style of play MUCH more attractive and chess-like than something like a Petrosian on steroids.
Disappointing!
If anyone knows of a Leela network which does NOT have that problem and has the FIGHTING EDGE in tactical melees rather than
always postponing gains for greater gains, but instead CONVERTING gains to substantive crushes, do let me know.
Cheerio.
http://www.pkoziol.cal24.pl/opental/dist11.zip
-
- Posts: 737
- Joined: Wed Mar 08, 2006 8:08 pm
- Location: Orange County California
- Full name: Stuart Cracraft
Re: Re:Talish Chess Engine?
@supersharp77 - Talish is a PRIVATE program whose code and binary will NEVER be shared.supersharp77 wrote: ↑Wed Apr 22, 2020 10:58 pm
Talish Chess Engine/Program?....No search results on a Chess Engine by the name of "Talish" but there is "OpenTal"