Faill Low nodes and move sorting.

[lauriet]

If you keep a list of fail low nodes (a bit like history tables) and use it to sort the worst moves last in the move ordering.......................
Is this the same (but opposite) of history sorting ?
Does it have any merit ?

Laurie.

[hgm]

One node's fail low is its parent's fail high.

I am not sure what you are proposing. History is not a list of nodes, but a list of moves. Nodes are in the TT. There it is already recorded whether a node failed low or high, by the bound type.

[Evert]

When you say "fail low node", do you mean "all node"?
In an all node you don't know which move is best (or worst), all you know is that none of them are good enough to beat alpha.

The idea of th history heuristic is indeed to store "good" moves before "bad" moves.

[PK]

There are two doable approaches to history scores

1) increment the counter on beta cutoffs and on new pv moves
2) increment the counter on beta cutoffs and on new pv moves and then decrement all the moves that have been tried in the same node, a la Fruit.

In both cases sooner or later you will need some code making sure that the values do not fall outside of bounds, i.e. divide them by 2 if they are too big/too small.

The first approach does what is intended: it sorts moves in a reasonable order.

The second one gives you additional magic number - move's history score. You can compare it to a constant and then decide that certain moves shouldn't be pruned etc.

[lauriet]

Sorry...thats about what i meant.....but after thinking more about it, it probly doesnt make any sense.

Laurie.