Changeset 2424
 Timestamp:
 Mar 6, 2019 9:40:32 PM (17 months ago)
 File:

 1 edited
Legend:
 Unmodified
 Added
 Removed

trunk/README.md
r2423 r2424 82 82 * [R and CLP  a quick start](https://cran.rproject.org/web/packages/clpAPI/vignettes/clpAPI.pdf) 83 83 * [Java and CLP  performs well](http://orinanobworld.blogspot.co.uk/2016/06/usingclpwithjava.html) 84 85 86 ## FAQ (from 2004)87 88 ### The barrier method sounds interesting, what are some of the details?89 90 The CLP barrier method solves convex QPs as well as LPs.91 In general, a barrier method requires implementation of the algorithm, as well as a fast Cholesky factorization.92 CLP provides the algorithm, and is expected to have a reasonable factorization implementation.93 However, the sparse factorization requires a good ordering algorithm, which the user is expected to provide (perhaps a better factorization code as well).94 95 ### Which Cholesky factorizations codes are supported by CLP's barrier method?96 97 The Cholesky interface is flexible enough so that a variety of Cholesky ordering and factorization codes can be used.98 Interfaces are provided to each of the following:99 * Anshul Gupta's WSSMP parallel enabled ordering and factorization code100 * Sivan Toledo's TAUCS parallel enabled factorization code (the package includes third party ordering codes)101 * University of Florida's Approximate Minimum Degree (AMD) ordering code (the CLP native factorization code is used with this ordering code)102 * CLP native code: very weak ordering but competitive nonparallel factorization103 * Fast dense factorization104 105 106 ### When will CLP have a good native ordering?107 The best outcome would be to have an existing ordering code available as part of the COINOR distribution under the EPL.108 However, if this is not possible, the native ordering will be made respectable.109 110 111 ### Is the barrier code as mature as the simplex code?112 The simplex code has been exposed to user testing for a while and the principal author, John Forrest, knows more about simplex algorithms than interior point algorithms, so the answer is "no".113 However, it performs well on test sets and seems to be more reliable than some commercially available codes (including OSL).114 115 116 ### Which algorithm should I use for quadratic programming and should I keep an eye open for any issues?117 The interior point algorithm for quadratic programming is much more elegant and normally much faster than the quadratic simplex code.118 Caution is suggested with the presolve as not all bugs have been found and squashed when a quadratic objective is used.119 One may wish to switch off the crossover to a basic feasible solution as the simplex code can be slow.120 The sequential linear code is useful as a "crash" to the simplex code; its convergence is poor but, say, 100 iterations could set up the problem well for the simplex code.
Note: See TracChangeset
for help on using the changeset viewer.