Engine strength is much worse when playing Chess960/FRC

[KhepriChess]

I decided to go off on a tangent, away from just adding more to improve engine strength, and added Chess960 support. I actually managed to get it all working, no perft errors and no errors when playing games.

The only issue now is that the strength seems to be considerably worse. My testing conditions aren't exactly ideal due to the lack of engines in the 1700-1900 Elo range that can play FRC. So I opted for just Stockfish set to Skill Level 3. In standard games (TC 10+0.1s), Khepri has around a 55-60% win rate. But in FRC games, the same time control, it's barely above 40%. It seems to fair slightly better with longer time controls, but still not nearly as strong as when playing standard games.

I suspect this is solely due to my evaluation and relying heavily on PSTs, which are tuned for standard with no consideration for FRC. But I'm not sure how to go about addressing that. Should I try FRC-dedicated tables (and what would those even look like)? Anyone else run into anything like this?

[KhepriChess]

I looked through the previous FRC tournament here and grabbed an engine (Pulsar) that was suitable to test against. Clearly not just a SF-testing issue. Ran two SPRT tests, one of standard games and one FRC games. Quite a clear difference between the two...

Standard: 140.9 +/- 44.6 Elo
Chess 960: -15.1 +/- 16.1 Elo

I did try making my existing PST symmetrical, sort of, by just grabbing the mirrored values if the piece's file was greater than 4. But that really made no difference.

[Modern Times]

Bagatur performs a few hundred Elo worse at chess960 than standard chess, so you're not alone !