Not necessarily. Infact if you do so it will probably be much larger than expected perft. Please note that perft is an _average_ estimate so you should give more weight to those nodes representative to the perft tree. Maybe the mode is candidate. But perft is different from tree search because you are not trying to maximize or minimize anything (just get average). For UCT tree search you bias the tree in such a way winning branches are explored more. Estimating perft starting from ply 3 already reduces the variance significantly. For the aggressive LMR tree we experimented with may be a depth of (perftD / 2) is more appropriate.That was a very drastic LMR as I recall. According to my formula the first move should get almost all the weight.
There is an exploration-exploitation dilemma as in the case of UCT. And if you want to use sigma for move selection, you should keep statistic on a tree. And I am not even clear how to do this for perft. UCT uses
Code: Select all
UCTValue(parent, n) = winrate + sqrt((ln(parent.visits))/(5*n.nodevisits))