Adding exactly k cuts at a root node.

[ComboProblem]

I would like to make sure for any root node that exactly k cuts are added to the root node. I have a general understanding of enabling adding cuts only at the root node from #763 and other related issues.

What I'm struggling with is how enforce "use exactly k cuts" condition. Is it possible to enforce such a condition within SCIP? If so, how does one go about doing this?

[Joao-Dionisio]

Hey @ComboProblem ! You probably want to use model.setParam("separating/maxcutsroot", k). Although this doesn't say use exactly k cuts, but rather at most k cuts. Would that be enough? For exactly k cuts, one would need to force SCIP to keep separating until that point. For this, I would play around with model.setParam("separating/maxcutsgenfactor", -1), but I'm not too sure on this one.

[ComboProblem]

Thank you for the reply. Either at most or exactly k cuts is fine. When I've tried say k=1 and disabling other cuts; I've found nodes past the root node with more than 1 cut applied. I am using the following example

from pyscipopt import Model, SCIP_EVENTTYPE, SCIP_RESULT, Eventhdlr, SCIP_PARAMSETTING

class checkCutsAdded(Eventhdlr):
    def __init__(self, model):
        Eventhdlr.__init__(model)
        self.count = 0 
        
    def eventinit(self):
        self.model.catchEvent(SCIP_EVENTTYPE.LPSOLVED, self)

    def eventexec(self, event):  
        if self.model.getCurrentNode().getDepth() > 1: # not at the root; we should have (afik) at most 1 cut added to this node
            self.count += 1
            self.model.writeMIP(f"node_lp_{self.count}.lp")
        if self.count == 10: # lets not write too many things
            self.model.interruptSolve()

class disableCuts(Eventhdlr):
    def __init__(self, model):
        self.model = model
    def eventinit(self):
        self.model.catchEvent(SCIP_EVENTTYPE.NODEFOCUSED, self) # the event is called whenever a node is about to be solved

    def eventexec(self, event):
        if self.model.getCurrentNode().getDepth() > 1: # if we aren't in the root node
            self.model.setSeparating(SCIP_PARAMSETTING.OFF) # disable separators
        else:
            self.model.setSeparating(SCIP_PARAMSETTING.DEFAULT)
            self.model.readParams("scip_disable_other_cuts.set") # see link  https://github.com/ComboProblem/parametricCutGeneration/blob/main/src/Experiments/paramFiles/scip_disable_other_cuts.set

model = random_mip_1(False) # from; not copy and pasting and afik not in the name space; https://github.com/scipopt/PySCIPOpt/blob/master/tests/helpers/utils.py
sepa = GMI() # From the example of seperator; not copying and pasting ; https://pyscipopt.readthedocs.io/en/latest/tutorials/separator.html
model.includeSepa(sepa, "gmi", "gmi test", priority=1000, freq=0)
model.readParams("scip_disable_other_cuts.set") # see link; https://github.com/ComboProblem/parametricCutGeneration/blob/main/src/Experiments/paramFiles/scip_disable_other_cuts.set
model.setParam("separating/maxcutsroot", 1) # only one cut at the root
model.setParam("branching/random/priority", 20000) # as per suggestion in previous thread to remove effects of strong branching
check_cuts = checkCutsAdded(model)
disable = disableCuts(model)
model.includeEventhdlr(disable, "disable_other_cuts", "disable other cuts aside from the targeted cut at depths not at the root")
model.includeEventhdlr(check_cuts, "check_node_mips", "output node mips to verify ")
model.optimize()

The first output can be found here I would expect that there should only be one cut added with the gmi name - but there are 5 (the gmi cuts start at around line 16195). I'm not really sure what is going on here or what I should be doing differently.

[Joao-Dionisio]

I think the issue is that maxcutsroot is about the number of cuts per round, and so we should also limit the number of rounds.

model.setParam("separating/maxcutsroot", 1)
model.setParam("separating/maxroundsroot", 1)                                                                                                                                                                 

[ComboProblem]

When I tried this change, it doesn't seem like that has changed the number of cuts present with the same test problem.