What about QSearch? I assume it just doesn't consider drops at all?
Indeed. I never managed to get a non-exploding QSearch that searched captures plus drops, even when I limited it to safe check drops. Even an extension for check evasions can be dangerous in QS. At least, if you allow drops as evasions. (I forgot to mention that: when in check at d=1, it is of course pointless to try check drops, and I try drop evasions instead, to prevent seeing false checkmates.) The problem is that the drop evasions can add captured pieces back to the board for the side that is being checked, and if the players alternately check each other this can go on indefinitely. (I never saw that problem in my Shogi engines, but in Crazyhouse it did happen.) So it is probably safest to ignore checks beyond a certain QS depth, or limit the number of checks, or limit checking to one player only. Very powerful pieces can cause problems with search explosion in QS anyway.
That is really painful. I wonder if it is possible to collect this information incrementally during the search rather than searching deeper to determine?
That is hard because of alpha-beta pruning, which might prevent the relevant moves from being searched. And you wouldn't want to do anything extra on the off chance that the current position might later run into a repetition; repetitions are not all that numerous in the tree.
As to castling: In Joker I had the problem that the castling bonus was so large that it could give rise to "positional horizon effect": it started to sacrifice material just to push the castling over the horizon. Which, predictably, only brought temporary relief for a permanent sacrifice. So in KingSlayer I make the associated King Safety bonus available in small steps. Clearing the path for castling to a location with a good Pawn shield already gives you part of the bonus for that shield. Or, in other words, the value of the castling right is a fraction of the bonus,and the more pieces are still in the way, the smaller the multiplier gets. And the availability of a second castling right gets added for 1/8 to that for the primary one. The Pawn-shield bonuses for all 4 locations are stored in the Pawn hash table, and also depends on the opponent having half-open files aimed at your shield.
Indeed. I never managed to get a non-exploding QSearch that searched captures plus drops, even when I limited it to safe check drops. Even an extension for check evasions can be dangerous in QS. At least, if you allow drops as evasions. (I forgot to mention that: when in check at d=1, it is of course pointless to try check drops, and I try drop evasions instead, to prevent seeing false checkmates.) The problem is that the drop evasions can add captured pieces back to the board for the side that is being checked, and if the players alternately check each other this can go on indefinitely. (I never saw that problem in my Shogi engines, but in Crazyhouse it did happen.) So it is probably safest to ignore checks beyond a certain QS depth, or limit the number of checks, or limit checking to one player only. Very powerful pieces can cause problems with search explosion in QS anyway.
That is hard because of alpha-beta pruning, which might prevent the relevant moves from being searched. And you wouldn't want to do anything extra on the off chance that the current position might later run into a repetition; repetitions are not all that numerous in the tree.
As to castling: In Joker I had the problem that the castling bonus was so large that it could give rise to "positional horizon effect": it started to sacrifice material just to push the castling over the horizon. Which, predictably, only brought temporary relief for a permanent sacrifice. So in KingSlayer I make the associated King Safety bonus available in small steps. Clearing the path for castling to a location with a good Pawn shield already gives you part of the bonus for that shield. Or, in other words, the value of the castling right is a fraction of the bonus,and the more pieces are still in the way, the smaller the multiplier gets. And the availability of a second castling right gets added for 1/8 to that for the primary one. The Pawn-shield bonuses for all 4 locations are stored in the Pawn hash table, and also depends on the opponent having half-open files aimed at your shield.