Thank you for reading my post and or the suggestion. I've attached the trace file as well as a more detailed one. FWIW, to show how similar the solution values are, here is the alglib solution. Please excuse the many decimal places. I generally ignore Visual Studio C++ numerical places beyond 1e-13 but I value extreme precision in my problem. Alglib filter sqp: Unit-Canonical Ellipsoid 1: (-0.329176838420615, 0.94401746027854, 0.0217633576606146) Unit-Canonical Ellipsoid 2: (0.930895021061298, 0.329232641233283, -0.158241359036905) In general space, this corresponds to General Ellipsoid 1: (-10.5967582185939, 206.347913283321, 6.77869831250489) General Ellipsoid 2: (18.8893523034911, -49.9697297257856, 570.805604711329) The Euclidean norm is: 620.237467405261 In code, of the 6 KKT condition equations, the worst feasibility is 9.7502106655156240e-06 and in my Excel "implementation" of my geometric convergence algorithm, the KKT feasibility evaluates as 0.0001294260728173 or roughly 1e-4.
In contrast, the point I find in my Excel "algorithm" differs from the ALGLIB solution by at most e-9 in unit-canonical space and e-8 in general space! My solution: Unit-Canonical Ellipsoid 1: (-0.32917684136241, 0.944017459191415, 0.0217633603207913) Unit-Canonical Ellipsoid 2: (0.930895019622935, 0.329232643072075, -0.158241363672689) In general space, this corresponds to General Ellipsoid 1: (-10.5967582026675, 206.347913213863, 6.77869828010731) General Ellipsoid 2: (18.8893522318179, -49.9697297427802, 570.805604707352) The Euclidean norm is: 620.237467405261 In code, of the 6 KKT condition equations, the worst feasibility is 9.0949470177292824e-12 and in my Excel "implementation" of my geometric convergence algorithm, the KKT feasibility evaluates as 0.00000000003547029 or roughly 3.5e-11.
Does this give you any hints? I realize it may seem like polishing an already very nice apple, but I'd love to squeeze out the extra accuracy from alglib if at all possible...
Attachments: |
File comment: trace_file("SQP.DETAILED,SQP.PROBING,PREC.F15", logDirectory + "trace.log");
trace_15.log [168.53 KiB]
Downloaded 1449 times
|
File comment: trace_file("SQP", logDirectory + "trace.log");
trace.log [47.92 KiB]
Downloaded 1514 times
|
|