Solvers

The base installation will take care of most of the numerical dependencies of TNO such as SciPy and NumPy. However a few additional solving libraries must be added to the environment.

From the default installation the SLSQP solver is already available via SciPy. At least two other packages need to be added: IPOPT and SDPT3.

IPOPT

IPOPT is a Python-package for Interior-Point-OPTimisation. The installation can be done running the following from the terminal:

conda install -c conda-forge cyipopt

More information about the solver and solver options are available here:

https://pypi.org/project/ipopt/

Loadpath solvers

MOSEK

MOSEK is a software package to perform convex optimisation. It is accessed through TNO via CVXPY. To use MOSEK you need to get yourself a license which can be purchased or obtained for free for scholars. See their website for more information on how to obtain such license and on their installation process see:

https://www.mosek.com/products/academic-licenses/

Once you have the licensing worked out we will install MOSEK and CVXPY python packages to use it.

The package CVXPY connects to several convex solvers to solve convex problems. You can learn about what it does through their website.

https://www.cvxpy.org/

We can install both CVXPY and MOSEK together with pip:

pip install cvxpy mosek

SDPT3

SDPT3 is a solver to perform conical convex optimisation. It is necessary to perform load-path optimisations. This solver is distributed on the MATLAB’s package CVX. Both need to be installed: MATLAB and CVX. The installation is broken in three steps:

Download and install MATLAB. Versions newer than R2022b are recommended.

https://www.mathworks.com/products/matlab.html

Download and install the CVX to MATLAB:

http://cvxr.com/cvx/download/

Install Python API for MATLAB to control it direclty from Python scripts:

https://www.mathworks.com/help/matlab/matlab_external/install-the-matlab-engine-for-python.html

Additional solvers

Additional solvers can be added to TNO as listed below:

PyOpt

Library with multiple solvers, some adequate for nonlinear optimisation. It has its own version of SLSQP and also other solvers including GA and DEVO. To install run:

conda install -c mutirri pyopt

More information abobut the library can be found here:

http://www.pyopt.org/download.html

PyTorch

The gradients and jacobians are done analytically in TNO. We add an option to compute these through auto-differentiation based on the package PyTorch. The following code should install it:

conda install pytorch torchvision -c pytorch