This package now exists in two versions: cobs and cobs99.

cobs99 (called cobs up to summer 2006) implements the algorithm of He and Ng (1999) for nonparametric quantile regression using constrained B-splines; more details are found on Pin Ng‘s COBS page.

The current package, cobs, is based on sparse matrices and the modified interior-point algorithm of Koenker and Ng (2005). A short paper, Ng and Maechler (2006), submitted for publication, explains and compares the old and new approaches. For more details, contact Martin Maechler, the package maintainer, or Pin Ng, the theory master mind behind cobs.

Reference manuals (on CRAN): cobs and (outdated, for reference): cobs99

Choosing the smoothing parameter lambda semi-automatically is currently only implemented using a version of SIC (aka BIC, the Schwartz’ or Bayes’ Information Criterion). That cannot be reliable for lambda->oo (since our version of SIC then tends to -oo). For this reason, low lambda values (< k) are greyed out.

In particular, look at the examples in ?plot.cobs, e.g., at plot(Sb1)

The non-simplex (but simplex-like) algorithm of the 1999 version can get into convergence problems in the case of ‘degenerate’ x-designs, e.g., if some x values are duplicated (multiple times). In such cases, the problem can often easily be remedied by using jitter(x) instead of x

Making better use of the internal sparse matrix algorithms, the code can be made considerably more efficient in the case where several lambda’s are used, e.g., for automatical lambda choice when lambda < 0.

Please, feel free to add! For bugs, the package maintainer would prefer e-mail containing a (small) R script with a reproducible example.