Hi,
after 20% progress in the test for the Super Constellation, as Ivan the Terrible would voice it: 'I think, I am onto something...'
First, a highlight about the hex score encoding method: the heavy weight byte relates to pawn-units whilst the light byte relates to the decimal fraction of the score.
Code:
e.g.:
&h8100-&h8000=&h0100 which is 256 in dec. base. Divided by 256 provides 1 pawn-unit as the score.
The figure "1" can be read directly on the right of the "8".
A 5 pawn-units score (one rook advantage) is &h8500.
Negative scores are just symetrical, -1 pawn-unit is &h7F00.
The light byte part of the score can range from &h00 to &hFF, thus 0 to 255, which is <1 once divided by 256,
therefore it stores the decimal part of the score.
Now, to the point: I noticed a blatant behavior discrepancy about the score management of the Super Constellation, compared to the Constellation Expert.
With the Expert, the score evolves in a very usual way, along with the computing time and depth reached: the light byte changes more frequently than the heavy byte. To apply a change to the heavy byte, the search needs to identify a change in the material balance, or to consider a large impact from positional change(s). Many variations within the search tree only provide small adjustments to the score, thru positional evaluation.
With the Super Constellation, the behavior follows another way: after a few seconds the score gets steady (due to the selection of the best move so far) and from then, the light byte value does not change any more. Along with the computing time and depth, the heavy byte may change (an in-depth capture being discovered), but the light weight byte is not any more changed by any positional concern. As far as I could check, it only may change along with a change in the material score, which I assume is a change in the root best move so far.
To my understanding, this reveals a PSH strictly applied either at root or at low depth in the search tree, providing a positional score per root move, followed by a material trading evaluation only for the rest of the tree search.
On another hand, the Constellation Expert is deemed to benefit from an improved positional play, and the fact is its score can be positionally refined much deeper within the search tree.
This is just all my understanding, but anyone can experiment in this field using the debug mode of the emulator; feel welcome to share your own outcomes!
Kind regards/MfG
Eric