I really should rename this thread "Texel Tuning for Lazy Dummies". Here's what has happened during about two weeks of occasional work on Myrddin and a lot of CPU time.
1) I added tapered eval, resulting in +29 elo. This was, of course, long overdue and took more work than any of the below steps.
2) I created a
very rudimentary Texel tuning system. This involved adding code to Myrddin to run through an EPD file and produce a static evaluation for each position. I used the ~1.4M quiet positions graciously provided by Alexandru Mosoi (Zurichess) as a training set and produced a base "score" (via the well-known sigmoid function) that I would attempt to improve via manual tuning. Yes, I did all this
manually, rarely tuning more than two variables at a time, and never more than three at a time - then moving on to the next pair (usually the mg/eg pair of a single factor).
3) As a first test, I simply started commenting out various bits of code in the eval to confirm that they would all result in a poorer score. Very surprisingly, commenting out
all of the pawn structure eval (passers, doubled pawns, isolated pawns, everything) resulted in an improvement to the score. As did removing the code to evaluate the pawn shield in front of a king. The only actual "tuning" that took place in this step was to decrease the bonus that was given for a piece attacking the square of the enemy king, or any square adjacent to the king. Reducing that bonus by 1cp resulted in a small improvement. I then ran a gauntlet (which takes about 18 hours) and got +31 elo, confirming that the system worked, and also that my pawn structure eval was complete crap.
Total improvement so far: +60 elo.
4) I then tuned mg/eg values for the bishop pair, rooks on open files, some of the pawn structure eval (doubled pawns, doubled pawns that are also blocked by an enemy pawn, and isolated pawns), and the pawn shield in front of a king. This resulted in another +26 elo, and a total improvement of +86 elo.
5) I then restored and tuned the passer code (including protected and connected passers, but removing code to evaluate rooks behind passers and passers that are outside the square of the pawn, as these two factors were not able to improve the score), mobility, and adding a "material adjustment" to the PSTs. I'm way too lazy to tune each value in the PSTs, but I thought I might be able to add a single mg/eg adjustment to the entire piece that would improve the score. This also worked out well. The result of all of this was another +23 elo and a total improvement of +109 elo. Goal achieved!
This entire process turned out to be surprisingly easy and shockingly effective, once I was finally able to get it through my thick and lazy brain how to do it. Thanks to Martin for the patient support, Graham and Lars for constantly asking me if I'm working on a new version, and (again) Alexandru for providing the training set.
So I'll be sending a version to Lars for independent testing. If anybody else wants to test an unreleased version, let me know.
