Videodr0me

I think the absolute simplest solution would be to allow two alternative parameters: current behaviour with a regexp string as parameter, and altenatively a list of parameter names where order...

I also had two related ideas about FPU - maybe its already been done, tested and discarded, but just to throw it in here. (1) if current search increases eval...

@jkiliani I read that discussion about Leela Zero Go and am not that convinced it applies to Leela Chess: (1) Go and Chess are different games. Reaching a terminal state...

@jkiliani I tested FPU in 10 tactical positions: using FPU prevented or prolonged leela from finding tactics. The higher FPU reductions the more pronounced was this effect. Skimming through the...

used defaults of 0.8 leela with net 258 just so see how fpu would affect tactical solution time.

defaults of 0.8 are: cfg_noise = false; cfg_randomize = false;

Sure they don't - the discussion you brought up from leela go was about FPU reductions and later morphed into a pruning single visit discussion. i just took it as...

@jkiliani No, i am not forcing every root move to have at least one visit. The adjustment only pertains to the accum vector, no visits are actually changed. Only the...

@killerducky I am pretty sure in proportion to visit count from the AZ paper actually means in proportion to non zero visit counts or laplacian adjusted visit counts. But you...

@jkiliani I am not saying anything is majorly wrong. Yes maybe the effect is mitigared by other noise - no problem. But as 0 visit nodes are real, and according...