GoldfarbIdnaniQuadraticSolver: possible bug?

[pabl...@gmail.com]

Hi everybody,
I posted a question about the GoldfarbIdnaniQuadraticSolver class a couple of days ago. After reading the original Goldfarb and Idnani paper I am wondering whether the version of the algorithm implemented within the Accord.NET library is working correctly.

According to the original paper, all the constraints should be active by the end of the algorithm. That is, the final solution should satisfy all the constraints.

However, that is not true for my particular problem. I am using Accord.Net to get a C# implementation of a prototype I previously implemented in Matlab, and I debugged all the code prior to the call to the Accord.NET classes' methods to be sure that I am feeding the algorithms with the correct parameters.

The algorithm finishes without displaying any error or throwing any exception, but many constraints are not satisfied. Is this a bug of the library's code?

[César]

Hi there,

Sorry for replying a little late. I saw the problem and was investigating what could be wrong. This issue had been reported before [1], and it is possible that this could have been a bug introduced after the translation of the original method from Fortran [2]. Please, may I ask if you could provide a sample of the function and the constraints so I could write more test cases for this situation? 

Best regards,

Cesar

[1]: https://code.google.com/p/accord/issues/detail?id=33

[2]: http://school.maths.uwa.edu.au/~berwin/software/quadprog.html

[pabl...@gmail.com]

Hi Cesar,
thank you for your help. Please find attached three cs files containing the sources of a sample problem. I also attached a csv file containing sample data.

[janaki...@gmail.com]

Cesar ,

Was there any new release to correct this? I am seeing problems with the GodfarbIdnani solver myself . In my problem I have 19 variables , 2 equality constraints , and 38 inequality constraints ( basically inequalities are that each of the 19 solution points should be in the closed set [0 1] ).

The code completed with a solution , but the last solution variable is set to a negative value ( violating one inequality constraint)

I downloaded the latest Accord.Net package , so I probably have one that is newer than the one mentioned in this post , so I am puzzled. It is possible that I have incorrectly specified the problem . What is an easy way to verify this? I can provide the code and variables if needed.

Thanks

Sri

[janaki...@gmail.com]

I scanned other posts and it looks like this is an open issue. Please ignore my own above . I'll spend some time to explore the issue further in the Accord.Net source for GoldfarbIdnani solver