Improved LP/MIP solver
On Sun, Dec 12, 2021 at 3:44 PM Julian Hall <jajhall at ed.ac.uk> wrote:
Dear All, I am leading the development of HiGHS, which is now the top performing open source linear optimization software on the industry standard benchmarks. In particular, our MIP solver out-performs SCIP, and is way ahead of the COIN-OR solver Cbc. HiGHS solves LPs via simplex or interior point, MIPs via branch-and-cut, and QPs via an active set method. We were wondering what interest there would be in developing an R interface to HiGHS. I'm not an R user, but have done a bit of searching and see references to Rsymphony and an interface to Lpsolve. Performance-wise Lpsolve is very poor, but I know that it has a community of devoted followers. I've not seen benchmark results for Symphony, but I know that Cbc is the preferred COIN-OR MIP solver when it comes to general performance. And, as I observed, the performance of HiGHS is way better than Cbc. Are people in the R community tearing their hair out over the performance of software requiring the solution of LPs or MIPs? Would a significantly better LP/MIP solver be valuable to the R community? Thanks, Julian -- Dr. J. A. Julian Hall, Reader, School of Mathematics, University of Edinburgh, James Clerk Maxwell Building, Peter Guthrie Tait Road, EDINBURGH, EH9 3FD, UK. Room: 5418 Phone: [+44](131) 650 5075 Email: J.A.J.Hall at ed.ac.uk<mailto:J.A.J.Hall at ed.ac.uk> Web: https://www.maths.ed.ac.uk/school-of-mathematics/people/a-z?person=47 [HiGHS]<http://www.highs.dev> My working hours may not be your working hours. Do not feel pressure to reply to this email outside your working hours. The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336. Is e buidheann carthannais a th? ann an Oilthigh Dh?n ?ideann, cl?raichte an Alba, ?ireamh cl?raidh SC005336.
Hello, Julian. I cannot speak for the R community, but as someone who needs optimization on a regular basis, this sounds intriguing. The fact that HiGHS appears to be FLOSS, and thus usable as-is in the corporate setting, appeals to those of us who use R in industry. Would you have any statistics on how the solvers in HiGHS compare with similar ones currently available in R, specifically the following in NLOPT [1] (which is called through nloptr): SLSQP (gradient-based) and COBYLA (gradient-free) both of which support equality and inequality constraints, and MMA/CCSA (gradient based) which supports inequality constraints? As for integer or mixed integer programming, I believe that there is a lot of room for improvement in R. Personally, I've resorted to using DEOptim with the "fnMap" entry calling a round function similar to [2]. So speaking for myself, giving richer options for optimization is a good thing, especially if the installation procedure can be simplified! Thank you, Avi [1] https://nlopt.readthedocs.io/en/latest/NLopt_Algorithms/ [2] https://stackoverflow.com/questions/42197353/how-to-set-integer-constraint-using-fnmap-in-deoptim-r