TalkChess.com

[Mark Pearce]

I've been drawn back into the computer chess community by the Rybka controversy, and decided to resurrect my old Fortran/assembly chess engine and update it with modern techniques. My weapon of choice is now C#

I put together a magic bitboard board representation, move generator, and move make/unmake. To my pleasant surprise, perft from the initial position not only produces the correct moves, but also runs at around 16M nodes per second on a single physical core. Modern hardware is just so impressive.

By "node", I mean generating the pseudo-legal moves, checking for move validity, then doing the move make/unmake. There is no bulk-counting or transposition table.

Flushed with success, I started to implement the evaluation function. Currently this is just the material score and interpolated PSTs using the opening/endgame trick popularised by Fruit and Toga.

Unfortunately, doing this evaluation on every leaf node in perft 6 reduced the nodes-per-second from 16M to 4M - ouch! Tonight I implemented an incremental PST evaluation, where the PST score is adjusted during every move make/unmake. This helped dramatically, bringing the node count back up to 12M.

So I have 2 questions.

Do the majority of people do incremental (PST or other) evaluation, given this magnitude of performance difference, or am I missing something?

My PST calcs use opening/endgame interpolation. Is this rather expensive (managed arrays?) trick really worthwhile from the ELO perspective, or is it better to stick with static PST values?
_________________
Theory is when you know something, but it doesn't work. Practice is when something works, but you don't know why. Chess programmers combine theory and practice: Nothing works and they don't know why.