About this release :
- This is primarily a strength increase release, mostly from experimenting with learning & eval features. (Also a few minor bug fixes.)
- On the plus side, this should be the largest strength increase between versions so far. Scored about +100 elo in 60 + .5 self-play. (Might be less against other engines or at different time controls.)
- On the minus side most everything else remains on my TODO list
I started experimenting more with learning & texel-tuning after taking a break from 1.9 where I had found a first pass of tuning effective. In the beginning it went better than I expected so I got pretty into it, then it started to go slowly, and I decided the time investment was getting too high again and I will probably take another longer break.
Specific Engine Changes from the ReadMe.txt :
More extensive work in evaluation tuning. Did Multiple passes and generated additional training games against a variety of strong engines to add to scored position list.
Made a fairly complete multi-threaded eval tuner with GUI window (unreleased for now since it requires a final step of pasting into source code.)
Added more eval features. Some specifics listed below. (Removed a couple too but mostly adding.)
Bishop diagonal, penalizing blocked pawn or defended pawn, bonus if enemy rook is on diagonal if not blocked/defended pawn.
Rook move to open file bonus. Even more different rook file types now has 8 I think. (eg. behind mobile pawn, pawn blocked by piece, pawn blocked by pawn, in addition to open/half open.) Bonus if enemy queen on file.
Knight center move bonus. Knight outpost by rank, higher outpost move value.
Altered some details of pawn eval. More direct pawn connected bonus (more on open file), unconnected penalty split for open/closed file + normal/backward/isolated, separate doubled score if can be undoubled. Some blocked pawn tweaks. Dynamic passed pawn scoring has couple more terms.
Various small search changes and bugfixes
Some tweaks to use of hard-coded tactical threats for LMR/LMP/Q-Search, overall probably uses them a bit more aggressively than 1.9.
(I thought maybe they'd be useless now but were +15 elo after tweaking, even more than the +8 for 1.9, though that was early on and may have already changed.)
About this release :
- This is primarily a strength increase release, mostly from experimenting with learning & eval features. (Also a few minor bug fixes.)
- On the plus side, this should be the largest strength increase between versions so far. Scored about +100 elo in 60 + .5 self-play. (Might be less against other engines or at different time controls.)
Easily enough for it to break into the Top 10 on the FRC list.
I updated the download on the webpage, if you downloaded it already you should re-download. I did manage to mess up the UCI options and the hash re-sizing, now those should be fixed. (I will try to do more thorough tests sometime later today when I have more time.) Thanks for the bug reports everyone! And if you notice anything still broken let me know.
The strength increase went super quick for a while (got 30 elo in 1 day once I set up the eval tuner for everything, plus 1 day of testing.) And I got excited about it and started spending way too much time again, and of course further progress slowed down after the initial jump. And testing became more frustrating with smaller differences, one tuning was up 16 elo after 1000 games and ended up negative after 2000. I probably will write a bit about my experiences with tuning in technical discussions, not really original, bit was still interesting (to me at least.)
Hi Jonathan,
Thanks for the new version.
Changing hash setting using own GUI, noPop version, crashes (v. updated Mar 22, 2020 downloaded).
Changing hash setting editing engineGUI.dft and/or engineWB.dft works. So not a big issue.
Engine (noPop) crashes on closure using Arena. Not a big issue as well
Kind regards
Tibono