From a numerical perspective, large speedups are to be seen on some classes of ill-conditioned models. This is due to new preconditioning techniques applied to the subproblems solved by conjugate gradient method in the Knitro interior-point/barrier algorithms. Furthermore, Artelys Knitro 10.3 offers several improvements to the internal linear algebra and linear solvers used internally for better robustness and efficiency.

Exponential speedups on selected nonlinear least square instance
Exponential speedups on selected nonlinear least square instance

Additional features

• Significant improvements in large-scale least-squares models, solving the instances 10 times faster!
• Improvements in the “feasibility restoration phase” used when Knitro is struggling to get feasible allowing faster detection of infeasible models.
• Overall efficiency and robustness improvements on general nonlinear models as well as significant improvements in efficiency on models with integer variables.