Pyomo scaling with GLPK

Question

I have a (fairly) simple Pyomo model with 5 parameters and a set of size 48 (which represent time intervals). GLPK works absolutely fine if I use a particular data file:

# Data file 

param : n := 48;

param : E_demand := 
1 231.674545 
2 223.328638 
3 218.047274 
4 212.285910 
5 214.539544 
6 213.940455 
7 216.871637 
8 205.824183 
9 208.905001 

(this continues in a similar vein up to index 48 and 4 more parameters).

But if I use another (only slightly different) data file, the problem takes much longer to solve (from less than a second to more than 20 minutes, I wasn't bothered to find out how much longer). If I just change two of the parameters to about 1/3 of their value (like below), the problem takes longer to solve.

param : E_demand := 
1 76.464996 
2 69.815002 
3 71.355003 
4 75.004997 
5 72.360001 
6 71.065002 
7 70.669998 
8 71.809998 
9 72.309998 

I think the problem must be to do with scaling, since if I gradually replace the smaller values from one data file to the other the problem takes more time until it becomes cumbersomely slow. Is there a way of changing the glpk scaling using Pyomo? Would using a different solver potentially solve this problem?

Answer 1

For ConcreteModel, you could always implement some form of scaling during model construction by checking the parameter values and applying the relevant scaling factors in the model formulation.

A similar discussion on variable/constraint scaling can be found on the Pyomo issues page: https://github.com/Pyomo/pyomo/issues/219

Answer 2

You may try gurobi solver. From personal experience I would say gurobi is much better and stable than GLPK for big opt problems. And if your opt problem is LP, then it should not make any solving difference when you are changing your data file to 1/3 of the original ones. Since it is the same thing, e.g.:

def constraint(m, n):
    m.E_demand[n] <= m.x

Either the Demand[1] = 200 or 73 does not make any difference. But ofc, I do not really know your model, so it is just an assumption.

here is the gurobi solver: http://www.gurobi.com/, it is also free for academic purposes.