cut nodes

[flok]

Hi,

I've measured that the number of cut-nodes in my program is around 13%. This seems to be a bit low.

What could be causing this?


regards

[Evert]

I've measured that the number of cut-nodes in my program is around 13%. This seems to be a bit low.

Is it?
I have no idea what other programs get there, but I get 11% for d=15 from the starting position using Jazz.

What could be causing this?

I don't think anything you do in the search can really change the number of cut-nodes (short of not actually taking the cut-off). Obviously doing things to the search bounds affects it (if you search every node with a full window, you get no cut-nodes at all), and so does how you handle fail high/low in PVS and null-move, but other than that I think the only thing that would affect it is the evaluation function.

[abulmo2]

How did you count them?

For example, in Amoeba, I have the following count (on bratko-kopec test set at depth 16, but the numbers are consistent with other measurements, from other test sets or real games):

Code: Select all

αβFailHigh     1,713,171  70.7%
αβFailLow        697,351  28.8%
αβExact           11,159   0.5%
total          2,421,681	

However, here, the total is the inner nodes that run the αβ loop. Including leaf nodes & QS search, I have got a total of 16,040,313 nodes, so the the proportion of cut nodes (αβFailHigh) is now only 10.7%.

[flok]

Evert, Richard,

The statement that it was a bit low was from someone on the irc channel.

But from your numbers I understood that "the soup is not eaten as hot as it is served" and that 15% of my program may not be an indication for where the bad results come from.


regards

[Volker Annuss]

Arminius from startpos and depth 17:

Code: Select all

cut:  501060  19.9700%
all: 2005657  79.9367%
exact:  2341   0.0933%

Positions from qsearch were not counted. Positions with hash cutoff, null move cutoff, draw by repetition, mate and stalemate were also not counted.