I want to use SCIP to run a pure cutting planes algorithm on a MIP.
Setting limits/nodes = 1 and increasing the limit of the maximum separated cuts at the root separating/maxcutsroot = 2147483647 alone do not produce the expected behavior.
What is the correct parameter setup?
Self-answer for future reference. The solution is provided below.
Several mechanisms must all be disabled. A cut separation loop in SCIP is
Accordingly, the following parameters should be adjusted (replace setParam with the equivalent call on your environment).
Prevents branching.
setParam("limits/nodes", 1)
Separation stops after a few rounds (10 is default) without dual bound improvement.
setParam("separating/maxstallroundsroot", -1)
Default is 2000 cuts per round at the root.
setParam("separating/maxcutsroot", 2147483647)
Each separator has its own maxroundsroot, and the loop exits when all separators exhaust their own cap.
SCIP_SEPARATORS = [
"aggregation", "clique", "closecuts", "cmir", "convexproj",
"disjunctive", "dualbound", "eccuts", "flowcover", "gomory",
"impliedbounds", "intobj", "lagromory", "mcf", "mixing",
"oddcycle", "rapidlearning", "rlt", "strongcg", "zerohalf",
]
for sep in SCIP_SEPARATORS:
try:
setParam(f"separating/{sep}/maxroundsroot", -1)
setParam(f"separating/{sep}/maxsepacutsroot", 2147483647)
except Exception:
pass # separator absent in this SCIP build
SCIP discards cuts whose efficacy is below minefficacyroot or whose direction is too parallel to existing cuts via minorthoroot.
setParam("separating/minefficacyroot", 0.0)
setParam("separating/hybrid/minorthoroot", 0.0)
Besides the mentioned answer, please note that neither SCIP nor the rest of today's modern solvers would be used as a pure cutting-plane algorithm, for many implemented hidden staff that may not be turned off anyway. However, many internal features need to be switched off. Some of them were written in the previous answer, and you can add the following for having more control over the solver behaviour.
model.setPresolve(SCIP_PARAMSETTING.OFF)
model.setHeuristics(SCIP_PARAMSETTING.OFF)
model.setSeparating(SCIP_PARAMSETTING.AGGRESSIVE)
model.setLongintParam("limits/totalnodes" , 1)
model.setIntParam("separating/strongcg/freq", -1)
model.setIntParam("separating/gomory/freq", 20000)
model.setIntParam("separating/aggregation/freq", -1)
model.setIntParam("separating/mcf/freq", -1)
model.setIntParam("separating/closecuts/freq", -1)
model.setIntParam("separating/clique/freq", -1)
model.setIntParam("separating/zerohalf/freq", -1)
model.setIntParam("separating/rlt/freq", -1)
model.setIntParam("separating/flower/freq", -1)
model.setIntParam("branching/random/priority", 20000)
And if you would like to turn off the cuts entirely:
model.setParam("separating/maxcuts", 0)
model.setParam("separating/maxcuts", 0)
model.setParam("separating/maxcutsroot", 0)
model.setParam("separating/maxrounds", 0)
model.setParam("separating/maxroundsroot", 0)
