Using mobility from the last full width plies movegenetion in evaluation. Sounds risky as quiescence captures have no effect on it but it's for free speed-wise and works better than expected.
Tuning down king safety better as pieces come off.
Passed pawns are scored better. Less to prevent wrong piece sacs but much more when they reach 7th.
In search doing researches on reduced nodes that were predicted wrong. A bit of Elo and longer PV's though I don't like how it sits in my code so probably just temporary.
I once tried a technique I called 'static mobility', which in theory should be an improvement on just using the mobility from the last pre-QS node. This determined the mobility of each piece separately, and added the resulting eval bonus to the value of that piece. That way, if during QS pieces are captured, their mobility contribution disappears with them. Of course there are other mobility changes that you would still miss (e.g. unblocking).
hgm wrote:I once tried a technique I called 'static mobility', which in theory should be an improvement on just using the mobility from the last pre-QS node. This determined the mobility of each piece separately, and added the resulting eval bonus to the value of that piece. That way, if during QS pieces are captured, their mobility contribution disappears with them. Of course there are other mobility changes that you would still miss (e.g. unblocking).
Good idea! Are you using this now and if not why not?
Though the current bang for the buck implementation works surprisingly well. Was expecting a lot of search instability but hard to notice more than usual. No doubt it does some funky things in branches we rarely see but ofcourse deeper search results in search instability/more accurate evaluation anyway so I was ready to accept this as part of that as the netto result is better play.
Splitting up into separate pieces was something I wanted to try but to be able to score them differently. (so that I can also remove specific bishop mobility in eval and get even more speeeed) And recalculating mobility per captured piece AFTER captures but your idea is simpler.
