[Scip] variable stats
Gregor Hendel
hendel at zib.de
Tue Oct 28 17:19:41 CET 2014
Dear James,
I am afraid you found a function of SCIP where the documentation is
incomplete.
The branch statistics contain the following variable information;
prio: The branch priority of this variable. Only the branching
candidates with highest priority are branched on at a node
factor: The branching factor, which is a softer way to prefer certain
variables by multiplying their actual score function.
..
depth: The average branching depth of a variable
branchings up/down: The actual number of times this variable has been
branched on in the respective directions.
sb: the number of times that this variable has been subject to the
strong branching look-ahead method.
Inferences up/down: The average number of domain reductions to other
variables observed after branching on this variables.
Cutoffs up/down: The percentage of infeasible subproblems obtained after
branching on this variable in the respective direction
LP Gain down/up: The average unit gain observed in the objective
function of the child node LP relaxations after branching on this
variable in the respective direction. For the unit gain, you measure the
actual, nonnegative gain of the LP objective of the child node ccompared
to its parent node, and normalize it by the variable fractionality.
Example: A variable with LP relax solution value 0.3 has been branched
on and we (later, or directly via strong branching) observed a gain of
+3 in the LP objective of the down child node, this is a unit gain of 3
/ 0.3 = 10 in the down direction.
pscostcount down/up: How many LP unit gains have been observed for this
variable?
Note that the latter number can be quite different from the actual
number of branchings for various reasons, when the created child nodes
were pruned or cut off without observing an actual gain.
Kind regards,
Gregor
Am 26.10.2014 um 20:24 schrieb James Cussens:
> Hi folks,
>
> I don't understand all of the information that is provided when SCP
> prints out statistics on variables after solving. This sort of thing:
>
> variable prio factor down up depth down up
> sb down up down up down up down
> up
> t_I#0 0 1.0 5 6 77.2 9 28
> 0 3.6 25.4 0.0% 0.0% 62.9177 101.0077 0.0
> 0.0
> ....
>
> Could someone tell me where to find an explanation of this? Apologies
> if this has been dealt with before.
>
> James
>
More information about the Scip
mailing list