As long as no TB position has been reached on the board, the goai is to reach a winning one (and to avoid a losing one).Joerg Oster wrote:@Ronald: is there a special reason to not allow probing of TBs when the root position is a TB position? (e. g. in a 6-man root position, still probing 3, 4 and 5-men TBs)
Once a winning TB position has been reached on the board, the goal becomes to convert it into mate. This is different.
If you want to find mate, it is useless to know that two positions are both TB win. You need to determine which gets you to mate quicker.
In KRPPvKR, almost all moves will be winning. In particular all those useless rook moves that make no progress towards a win. How is probing them going to help?
If you want to probe just to get as fast as possible to a pawn move or capture, there is a better solution: just pick the root move that minimises DTZ. But this is going to give ugly play and far more complaints from users.
Normally SF's search can be relied upon to make progress. So that's what the current code relies on, with filtering of moves at the root (using DTZ tables) giving the absolute guarantee that even if SF's search gets confused, the game will be won before the 50-move rule kicks in.
Of course it is possible to put in more intelligence. If the engine keeps track of the evaluations and DTZ values at the root since the last pawn move or capture and sees that things are stagnating, it could decide to play only DTZ-optimal moves until the next pawn move or capture just to make some sort of progress quickly.
SF is already doing something like this: if it detects that a repetition has occurred (on the board) since the last pawn move or capture, it only plays DTZ-optimal moves. This ensures that the repeated position is not reached a third time (which would mean draw). But it is pretty sure that Marco will rip this out once he gets to it (try to explain to him that such things are useful), so actually adding intelligence is not really in the cards in case of SF.