Release notes

What’s new in Knitro 13.1 ?

  • Knitro 13.1 offers significant performance improvements when using the BFGS/LBFGS Hessian approximation.

  • Knitro 13.1 updates the Knitro-MATLAB interface to support the MATLAB “problem- based” API. Many new examples demonstrating the “problem-based” API are added to the knitromatlab examples directory.

  • Knitro 13.1 incorporates the Apple Acelerate BLAS by default (blasoption =4) on MacOS machines with M1 processors, resulting in faster performance on this platform.

  • Knitro 13.1 updates the initial point specification when using Knitro though the Julia interface. This may improve performance in cases where the user does not specify an initial point.

  • Knitro 13.1 adds a new API function KN_set_mip_var_primal_init_values() that can be used to set initial primal variable values for MIP. This point may be used to search for an initial feasible point and may be different from the starting point for the root relaxation subproblem.

  • Knitro 13.1 introduces new user options for control of finite-difference gradient computations. Option findiff_estnoise, performs an estimation of the noise in the problem that can be used to automatically set an appropriate finite-difference steplength. Option findiff_numthreads can be used to control the number of threads used for the finite-difference gradient computations, when these evaluations can be performed in parallel.

  • Knitro 13.1 introduces the new user option bar_mpec_heuristic, which controls a new heuristic for solving MPEC models when using the barrier/interior-point algorithm.

  • Knitro 13.1 introduces the new user option mip_restart, which controls restarts for the MIP branch-and-bound code.

  • Knitro 13.1 introduces the new user option mip_heuristic_misqp, which controls a new heuristic for the MIP branch-and-bound code.

Bug Fixes in Knitro 13.1.0

  • Fixed issue that could lead to non-deterministic behavior on mixed-integer models when applying mixed-integer rounding cuts and using more than one thread.

  • Fixed issue with parallel finite-difference gradients used with mixed-integer models in the branch-and-bound algorithm.