At the moment I am giving a try to a connected idea: I define history counters as living on those PV or PV-candidate nodes you mentioned.Aleks Peshkov wrote:Search order of PV-candidate nodes is important, so IMHO it is practical to keep searched subtree size counters as a merit of move importance not only for root positions, but for all positions in a PV. An array of about 40 moves wide x 20 plies deep is not a huge memory waste.
Meaning that they are updated only in those nodes.
My hope is, this should avoid tree search statistics which are too noisy, and if noise still occurs, I take it as an indicator that my engine is into deep water (well, murky at least...).
This is because I want to extract "positional" informations out of statistics (I view my engine more as a "computational tool").
Probably storing a more detailed information for every node (what you are advocating) could be a more complete description, I hope not too much.
Cheers, Mauro
PS BTW, the "reduced" history counters work well, for me