Work in Progress

A version of SDPpack written in C is in preparation. How much speedup this will gain over the present version is not yet clear, since use of mex files makes Matlab quite efficient.

Other possible algorithmic improvements have been investigated. These include:

• Variations on Mehrotra's predictor-corrector scheme, including higher-order corrections and other formulas for the choice of centering parameter. Typically these lead to a reduction in the number of iterations, but increased CPU time.
• Extension of Gondzio's multiple centrality corrections for LP to SQLP. This led to a reduction in computation time on the largest of the truss problems (truss8) discussed in Appendix C.
• A simplified homogeneous self-dual scheme, which in principle assists with detection of infeasibility. However, when used to solve feasible problems we found this scheme gave substantially less accurate solutions to the original problem than the present code, and furthermore did not in practice resolve the feasibility issue in cases where feasiblity is in doubt, for example the control LMI's in Appendix C.
• Choosing the starting points by solving two least squares problems, as is common in LP. This has not proved to be a good choice in our experience.

None of these show enough improvements to the present code to justify their inclusion in the package, but this may change in the future.