I tried to fix up the check thing last night and got another nice little improvement. I put detection of 'safe' checks into my move generation routine and these are now scored the same as killers, so they will go into the list above the point where the search might be aborted. After 5000 games today, it seems this change is +11 (+8,-8 at 95% conf) better than the previous best implementation. Still haven't tried the more traditional LMP version of pruning each move individually, but I will get around to that.diep wrote:Implementation technical you're correct, but that doesn't count as a 'new prunings idea' IMHO, as otherwise every different way to implement something also must be seen as a 'new algorithm', which is a simple way to total refute it as a new idea.Ferdy wrote:I guess your idea is new, I have seen others just doing move prunning only, but yours is shall I say late move iteration prunning because you are prunning the whole iteration.dchoman wrote:Thanks everyone for the reference to LMP. It indeed appears to be essentially the same idea. I do have one question. In LMP, does one make every move and then prune them each individually... or does the whole remainder of the movelist get pruned in one go after a certain point.... or do different implementations do it differently? The former (move by move) might be safer than what I was trying (remainder in one go). I was worrying about missing a good check or something late in the movelist (as another poster also pointed out) and the move by move approach would avoid that, although it would be slower.
Is there a preferred approach?
- Dan
I tried your idea at higher depths and very late moves in move list without even bothering the depth specific value and got encouraging result.
There are some positions in the search that even a check move would not be able to do anything, so this approach is the solution.
Note that i'd take the +30 elo.
Bit amazed he doesn't except for giving checks either.
For Diep any of this doesn't work currently, though i retest things like this every few years though.
A problem there is always that for example a move like Na6-c5, a simple shuffle move is going to improve the evaluation always more than what most pawncaptures add to the evaluation.
- Dan