Getting domain error because of scaling set to 0.

[yd2612]

Hi,

I am getting "ValueError: domain error" in misc.py  on  W['di'] = W['d']**-1 . Any thoughts? Thanks.

[Martin]

Do you have an example where this happens? Perhaps your problem is poorly scaled?

[yd2612]

Hi,

PFA the pickle file containing the problem parameters in a dictionary. Problem is in the following form (min c^Tx s.t Ax=b , lo<=x<=hi). It is fit2d problem from the netlib dataset. I am also attaching small python script which I am using to run the optimization using cvxopt. Please let me know if you need anything else. Thanks a lot. 

[Martin]

Your problem is poorly scaled: some elements in h have magnitude around 1e20 while most elements of h are between 0 and 10. 

[yd2612]

Hi,

Thanks, removing the redundant constraints worked(large upper bounds). But in a different problem I am getting "Terminated (singular KKT matrix)". From different post I found out that it could be scaling issue again, so I tried scaling A and G by l1 norm of rows, also tried diagonal sacling (where scale found out from the first run of the algorithm) but I am still getting the singular KKT problem. Any thoughts? I am attaching the problem and the script I am using(same format) in case you need them. Any help is greatly appreciated. Thanks.

[Martin]

This is a similar issue: the elements of h are either close to zero or close to 1e20. If some of these constraints are essentially vacuous upper bounds on some variables, it is better to leave them out altogether.

[yd2612]

Thanks. I was in the impression that I was setting entries of h also to zero but I was only setting entries of G to 0. Thanks a lot.