Oct 23, 2011

[Info] Slight improvement

It seems the bug fixes I mentioned in my previous post has given a slight but measurable improvement:

Score of M1-1 vs M1B: 420 - 366 - 238 [0.53] 1024

(M1-1 being the newest version with the fixes)

Fortunate since I hadn't committed changes to the svn repository for a while and it started to be hard to see exactly what was changed.

This run includes the tapered eval change. I did a 1024 games test run with only that and it turned out to be exactly 50/50, that is no change in strength whatsoever. I had expected a slight improvement from that, but I guess my eval isn't really tuned for it. But no change is just fine, since I now have it in there and can start building from it.

I've also added a whole bunch of testing features and rewritten my "console mode" quite a bit. For example it's now possible to define a file with fen-positions and do an "evalmirror" on all of them to spot bugs. Might not be so fun for the average user, but extremely useful for me.

Before I release 1.0 I'll make it possible to run testsets in the epd format, directly from within Mediocre.


I've tried desperately to improve on the passed pawn evaluation. As it stands it's not very good and I suspect there are a lot of improvements that can be done.

But all my attempts so far has been in vain, I've tried Ed Schröder's way of evaluating them, and also Stockfish's and Crafty's, and some combinations between them coupled with my own ideas.

But everything I do seem to severely hurt the playing strength. And with severely I mean things like 100 rating points worse or so.

I'll keep trying though, I'm sure I'll hit the magic implementation at some point.

No comments: