Staged move generation and killers

[hgm]

Most of the move generation is in quiescence, and making none of the moves there sounds strange. Or do you exclude high takes low captures e.g. via SEE? Of course, in QS, you only generate captures and not quiet moves (except maybe for check evasions).

Even if you don't prune bad captures, QS branches will typically terminate in leaves where all captures are futile. Nodes where there are no captures at all are of course a special case of that. But even there you would also know that after running a move generation.

[Ras]

But even there you would also know that after running a move generation.

Ah right, that makes sense because all the loops and stuff are still being run, only that there is no capture anymore to be added to the move list, and then of course no make/unmake either.

[Michael Sherwin]

One trick that I do in RomiChess that I got a lot of consternation over at the time was for the remaining moves Romi does a very shallow search of 1 or 3 additional plies and then plays the move with the best score first. No one believed that it would help so I modified TSCP1.81 to do the same and TSCP gained about a 1.5 ply deeper search in the same amount of time. Here is the code in RomiChess.

Code: Select all

    case ADDMOVES:
      h->phase = NEXTMOVE;
      AddMoves(h->node, depth);
      if(h->node == (h+1)->t) return FALSE;
      if(depth > 3) {
        for(node = h->node; node < (h+1)->t; node++) {
          Sort(node, (h+1)->t);
          MakeMove((moves *)&node->m);
          if(GetReps())
            node->score = 0;
          else {
            inShort++;
            node->score = -Search(-beta - 180, -alpha + 20, depth > 7 ? 3 : 1, 0);
            inShort--;
          }
          ClrReps();
          TakeBack();
        }
      }        
    case NEXTMOVE:
      if(h->node + 1 == (h+1)->t) {
        h->phase = NOMOVES;
        return TRUE;
      }
      Sort(h->node, (h+1)->t);
  }

And Sven, if you go to post negatively about this idea, Romi has asked Jumbo to sit on you!

[tmokonen]

It looks similar in concept to internal iterative deepening to me. Am I missing a subtlety here?

[hgm]

The difference is that he opens the window, so that he gets exact scores that he can sort on, rather than an upper bound that is meaningless for other purposes than singling out the best move.