Change objective function during optimization

Ongoing

Tobias Völk

• July 14, 2022 15:47

We cannot use a certain type of cut with Gurobi because it requires the replacement of a part of the objective function in the callback. It would be cool if there was some "on-your-own-risk"-way of doing this, maybe a parameter for enabling this and otherwise having Gurobi complain that this is a terrible idea?

0

• 

  Jaromił Najman

  Gurobi Staff

  • August 09, 2022 11:59
  • Edited

  Replacing parts of the objective function would mean that Gurobi has to process the change through the whole B&B tree thus far and possibly adjust cuts and all relevant datastructures, resulting in a significant workload. This may in worst case even mean to throw away the whole B&B tree progress. Thus, we currently do not plan to implement this feature.

  I understand, that in some cases it might be possible that the state of the tree may remain unchanged but this seems to be a very specific situation. You could try to terminate the solution process, save the solution point, add the cut, adjust the objective function, provide the saved solution point as a MIP start, and re-optimize. This might not 100% reflect your idea but it might be just good enough.

  Best regards, 
  Jaromił

  0
• 

  Cassiano Tavares

  • August 21, 2023 13:06

  

  

  Hi Jaromil, a have a similar problem.

  I need to delete and insert the same objective function during each iteration in the loop of the rolling horizon. I  coded my model in Pyomo framework and used the Gurobi Persistent to convert my model to Gurobopy. This documentation (https://pyomo.readthedocs.io/en/stable/advanced_topics/persistent_solvers.html ) states, "We can also add or remove variables, constraints, blocks, and objectives.". But the GurobiPersistent class hasn´t implemented these methods.

  My objective function is presented following:

  model.exprobj = pyo.Expression()

  if RestrEquilibrioVarietal == 'Soft1': 

      if RestrEquilibrioEtario == 'Soft1': 

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          - sum( PenEqVar[t] * sum(sum(ViolaUmin[v,l,t]+ViolaUmax[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          - sum( PenEqIdade[t] * sum(ViolaWmin[g,t]+ViolaWmax[g,t] for g in model.set_G) for t in model.set_T) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

      elif RestrEquilibrioEtario == 'Soft2':

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          - sum( PenEqVar[t] * sum(sum(ViolaUmin[v,l,t]+ViolaUmax[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          -PenEqIdade[0] * sum(ViolaWmin[g]+ViolaWmax[g] for g in model.set_G) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

      else:

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
          - sum( PenEqVar[t] * sum(sum(ViolaUmin[v,l,t]+ViolaUmax[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

  elif RestrEquilibrioVarietal == 'Soft2':

      if RestrEquilibrioEtario == 'Soft1': 

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          -PenEqVar[0] * sum(ViolaUmin[v]+ViolaUmax[v] for v in model.set_V) \
                          - sum( PenEqIdade[t] * sum(ViolaWmin[g,t]+ViolaWmax[g,t] for g in model.set_G) for t in model.set_T) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

      elif RestrEquilibrioEtario == 'Soft2':

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          -PenEqVar[0] * sum(ViolaUmin[v]+ViolaUmax[v] for v in model.set_V) \
                          -PenEqIdade[0] * sum(ViolaWmin[g]+ViolaWmax[g] for g in model.set_G) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

      else:

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          -PenEqVar[0] * sum(ViolaUmin[v]+ViolaUmax[v] for v in model.set_V) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )
  else:

      if RestrEquilibrioEtario == 'Soft1': 

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          - sum(PenEqIdade[t] * sum(ViolaWmin[g,t]+ViolaWmax[g,t] for g in model.set_G) for t in model.set_T) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

  
      elif RestrEquilibrioEtario == 'Soft2':

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          - PenEqIdade[0] * sum(ViolaWmin[g]+ViolaWmax[g] for g in model.set_G) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

      else:

          model.exprobj = sum(sum(sum(theta[v,l,t] for v in model.set_V) for l in model.set_L) for t in model.set_T) \
                          - sum( PenMaxVar[t] * sum( ViolaVmax[f,t] for f in model.set_F ) for t in model.set_T )

  model.obj = pyo.Objective(expr= model.exprobj, sense = pyo.maximize)

  opt = pyo.SolverFactory('gurobi_persistent')
  opt.set_instance(model)

  In each iteration of the rolling horizon loop, I need to delete and insert the same model.obj (objective function).

  How I can implement this?

   

   

   

   

  0

Please sign in to leave a comment.