💡📝H. G. Muller wrote on Mon, Jun 29, 2020 09:04 PM UTC:
I now put a somewhat more refined search in the AI of the Interactive Diagram, in the hope that this would make it more human-like w.r.t. the things it will overlook. It now focuses its attention more on moves that are logical continuations of what happened before. In particular moves that 'touch' squares that were mutated one or two ply earlier. That would include recaptures, plunder raids, discovered threats, pin punishment. But also hopper activation; it doesn't matter whether the mutation was evacuation or placement; if a move goes over or to it, it gives it more attention than other moves.
As this search is much less fixed-depth than the previous one, the meaning of the ply setting has become a bit vague. If we count PV moves as one ply, most non-captures elsewhere in the tree would count double. I made the '2 ply' setting such that it will still search every reply to the move it intends to play (validated by an exchange evaluation), which means that the PV will actually be 3 ply long. Recaptures are now only not counted in the depth when a lower-valued piece captures a higher or an unprotected one; recapturing a lower protected piece doesn't get more attention than any other capture (or PV move).
It is still true that at 2 ply it wouldn't see a (new) mate-in-1 threat against it, and at 3 ply it will. It would see it when it can checkmate you in one at 2 ply.
As this is a bit more bug-prone than the very simple original search, please let me know if it shows erratic behavior. I am aware that its play with Pawns currently sucks. I could easily put in some evaluation that would solve this for FIDE/Shatranj Pawns, but that would break the generic nature of the AI. It does seem to be the kind of Pawn for which this is most important, though; there don't seem to be many subtleties with Asian or Berolina Pawns. (And of course the overwhelming majority of CVs have FIDE Pawns.)
I am open for suggestions for what additional features (other than pieces with nearly arbitrary moves) could be supported by the AI to make it more useful.
I now put a somewhat more refined search in the AI of the Interactive Diagram, in the hope that this would make it more human-like w.r.t. the things it will overlook. It now focuses its attention more on moves that are logical continuations of what happened before. In particular moves that 'touch' squares that were mutated one or two ply earlier. That would include recaptures, plunder raids, discovered threats, pin punishment. But also hopper activation; it doesn't matter whether the mutation was evacuation or placement; if a move goes over or to it, it gives it more attention than other moves.
As this search is much less fixed-depth than the previous one, the meaning of the ply setting has become a bit vague. If we count PV moves as one ply, most non-captures elsewhere in the tree would count double. I made the '2 ply' setting such that it will still search every reply to the move it intends to play (validated by an exchange evaluation), which means that the PV will actually be 3 ply long. Recaptures are now only not counted in the depth when a lower-valued piece captures a higher or an unprotected one; recapturing a lower protected piece doesn't get more attention than any other capture (or PV move).
It is still true that at 2 ply it wouldn't see a (new) mate-in-1 threat against it, and at 3 ply it will. It would see it when it can checkmate you in one at 2 ply.
As this is a bit more bug-prone than the very simple original search, please let me know if it shows erratic behavior. I am aware that its play with Pawns currently sucks. I could easily put in some evaluation that would solve this for FIDE/Shatranj Pawns, but that would break the generic nature of the AI. It does seem to be the kind of Pawn for which this is most important, though; there don't seem to be many subtleties with Asian or Berolina Pawns. (And of course the overwhelming majority of CVs have FIDE Pawns.)
I am open for suggestions for what additional features (other than pieces with nearly arbitrary moves) could be supported by the AI to make it more useful.