Implementing multiple branching rules in the same branch and bound tree in PySCIPOpt

Question

I would like to implement a custom branching rule initially (for a few nodes in the top of the tree), and then use Scip's implementation of vanilla full strong branching rule (or some other rule like pseudocost). Is this possible to do using/by extending PySCIPOpt?

import pyscipopt as scip
import random

class oddevenbranch(scip.Branchrule):

def branchexeclp(self, allowaddcons):
    '''
    This rule uses the branching rule I have defined if the node number is odd, 
    and should use strong branching otherwise.
    '''
    node_ = self.model.getCurrentNode()
    num = node_.getNumber()
    if num % 2 == 1:
        candidate_vars, *_ = self.model.getLPBranchCands()
        branch_var_idx = random.randint(0,len(candidate_vars)-1)
        branch_var = candidate_vars[branch_var_idx]
        self.model.branchVar(branch_var)
        result = scip.SCIP_RESULT.BRANCHED
        return {"result": result}

    else:
        print(num, ': Did not branch')
        result = scip.SCIP_RESULT.DIDNOTRUN
        return {"result": result}

if __name__ == "__main__":

    m1 = scip.Model()
    m1.readProblem('xyz.mps') # Used to read the instance
    m1.setIntParam('branching/fullstrong/priority', 11000)

    branchrule = oddevenbranch()
    m1.includeBranchrule(branchrule=branchrule,
                        name="CustomRand", # name of the branching rule
                        desc="",           # description of the branching rule
                        priority=100000,   # priority: set to this to make it default
                        maxdepth=-1,       # maximum depth up to which it will be used, or -1 for no restriction
                        maxbounddist=1)    # maximal relative distance from current node's dual bound to primal 

    m1.optimize()

I wonder what is causing this behavior. Does it need to call branching multiple times to perform strong branching?

2 : Did not branch
2 : Did not branch
2 : Did not branch
2 : Did not branch
2 : Did not branch
2 : Did not branch
2 : Did not branch
2 : Did not branch
2 : Did not branch
2 : Did not branch

Answer 1

I believe you can achieve this effect. SCIP has several branching rules and executes them one by one according to their priorities until one of them produces a result.

The default rule "relpscost" has a priority of 10000, so you should create a custom rule with a higher priority.

If you do not want to use your own rule at a node (deeper in the tree), you can decide in the branchexeclp(allowedcons) callback of your rule to return a dict

{'result': pyscipopt.SCIP_RESULT.DIDNOTRUN}

which should inform SCIP that your rule did not execute and should be skipped, in which case the "relpscost" rule should take over (see here).

Edit: It is not quite obvious to me how your instance looks like so I am guessing a bit here: I think you are right in assuming that the multiple calls at node 2 are due to the strong branching. You can check by switching to another backup strategy, such as "mostinf".

Additionally, it is really helpful to check out the statistics by calling model.printStatistics() after the optimization has finished. I tried your code on the "cod105" instance of the MIPlIB benchmarks and similarly found a set of "Did not branch" outputs. The (abbreviated) statistics regarding branching rules are the following:

Branching Rules    :   ExecTime  SetupTime   BranchLP  BranchExt   BranchPS    Cutoffs    DomReds       Cuts      Conss   Children
  CustomRand       :       0.00       0.00          1          0          0          0          0          0          0          2
  fullstrong       :       5.32       0.00          7          0          0          0          7          0          3          0

So the fallback rule is in fact called several times. Apparently, the "fullstrong" rule calls "CustomRand" internally, though this is not reflected in the actual statistics...