Besides any misundertanding on our "disputes", only one of us is right; but it is not really important as we are not vying for the Nobel award! Or to head ICGA!
We all learn otherwise how are we going to keep up with Robert Houdini who has the advantage of magic tricks! Through this thread, I now learned that leave nodes are not that "far away" from the root. And people apply the term 'ply' as in plywood as well to half-moves. And amplification ... OPPS...
Sven Schüle wrote:
Chan Rasjid wrote:It was you who explained how a TT entry of infi -5 may cause its grandparent node to fail high and hash an amplified score of infi - 7. This first amplification is the key.
No! This is no "amplification", it is standard backing up of tree values. A score "mate in 5 plies from here" at node A1 is backed up as "mate in 7 plies from grandparent of A1" to the grandparent (if the line is forced).
Chan Rasjid wrote:The point is that the position of the grandparent node is not identical (a transposition) to that of its grandchild - two different positions. But this grandparent position may be reached via many different paths (transposition). When they are reached, further amplification occurs. If reached at ply 14, it would fail high with infi - 7 - 14; and this further amplified score gets propagated down towards the root again. And such processes keeps repeating.
Alpha-Beta with mate-distance pruning prevents going deeper than the shortest mate you already found from a given node, so you never really get there. But in my opinion even normal alpha-beta without mate-distance pruning will prevent the case you mentioned.
This was what you wrote earlier :
4) The grandparent of A1, which is at ply=10, gets back "+(infi-17)", fails high, and stores "+(infi-7)" as lower bound in the TT, i.e. "mate in at most 7 plies from here".
You said "stores" which means (infi - 7), an amplification, was hashed in the TT. The originating node hashed only "infi - 5". In Natale's case, the fail high reached the root node and what was hashed was (infi - 5 - 12 + 0)!
I have not examined about the part how alpha-beta would prevent any longer mates being propagated down. Off the hat now, the forced lines may stop after a few plies down and the black king may have a score within bounds and fail high stops. So as alpha-beta rewinds further down, the best score may again be a non-mate value. But the problem again is with the nodes higher up from this node - there are so many transpositions(KNB-K) and many have amplified TT scores hashed. So longer mates starts propagating down and search may start to fail high again with amplified mate scores. The processes repeats with each root iteration.
The thing is not many others join in this thread to express their humble opinion except the ICGA accredited experts Sven and Rasjid - and they are in disagreement!
Don't believe when you're told "There's no free lunch!" There is Linux.