Saturday, May 4, 2013

To Null Move or not to Null Move

 
After tuning the evaluation weights I'm currently using a PBIL based evolution to tune the parameters within the search framework of iCE. It includes control parameters for Extensions, Futility Pruning, LMR, IID, Delta Pruning, Lazy Eval and of course Null Move.

Part of the genom is a bit for most features that allows the evolution to disable a feature or restrict it to PV or None PV nodes.




The evolution is still running but is already pretty converged. During the course of the run I realized that the evolution is unsure whether the Null Move should be performed always or restricted to Non PV nodes. This comes a bit as a surprise because I never considered null moving in a PV node even theoretical sound (I just included the bit in the genom for completeness).

I checked a few other engines including stockfish and crafty and although crafty is using null move extensively it excludes it in PV nodes. The only engine that seems to do it is Texel


In a PVS framework PV nodes are rather rare, so it might make little difference whether Null Moves are used here or not. Maybe this is the reason the evolution has a hard time to converge this bit. But I'm still wondering whether it breaks something if used. A PV that contains illegal moves is surely questionable.

If the final genom has it on, I will probably have to spend some additional asserts in the framework.

How are others approaching that ?

Followup: 

I performed a test of the final genom with and without null moving in PV and it made almost no difference. Therefore the evolution was not able to converge that. The one that used null move in PV nodes performed 4 ELO better but well within the error bar. So for the moment I will disable it.

No comments:

Post a Comment