Skip to content

NOTES.md

Latest commit

 

History

History
131 lines (94 loc) · 4.6 KB

NOTES.md

File metadata and controls

131 lines (94 loc) · 4.6 KB

Developer notes for OptimPack

Publishing a new release

To prepare a new public release, follow these steps:

  1. If the API has changed, update the interface version number in macros libopk_version, libcobyla_version, libbobyqa_version, and libnewuoa_version libopk_la_LDFLAGS in configure.ac. See libtool documentation for explanations. The interface version number is c:r:a (for current, revision and age), revisions between c and c-a are supposed to be supported by the library. If the library source code has changed at all since the last update, then increment revision (c:r:a becomes c:r+1:a). If any interfaces have been added, removed, or changed since the last update, increment current and set revision to 0 (c:r:a becomes c+1:0:0). If any interfaces have been added since the last public release, then increment age. If any interfaces have been removed or changed since the last public release, then set age to 0. Never try to set the interface numbers so that they correspond to the release number of the package.

  2. Bump release numbers in README.md and in the macro project_version in file configure.ac.

  3. Commit the changes made in steps 1 and 2.

  4. Update the CHANGES.md file and commit the changes.

  5. Rebuild configure script and make the archive:

    cd build
make
make dist-gzip
mv optimpack-${VERSION}.tar.gz ../releases/.

  6. Push the changes on GitHub:

    git push

  7. Open OptimPack page at GitHub and click on the releases tab and then on the Draft a new release button.

    • Tag the version.
    • Add a brief description and a detailed description of the most important changes.
    • Check/uncheck This is a pre-release.
    • Attach the release archive.
    • Click on the green Publish release button.

  8. In the local git repository, pull the tag with:

    git pull --all

To have binaries automatically build by Yggdrasil:

  1. Fork the Yggdrasil project on GitHub:

  2. Clone the fork, make a branch, and edit the file Yggdrasil/O/OptimPack/build_tarballs.jl:

    git clone git@github.com:emmt/Yggdrasil.git
cd Yggdrasil
git co -b optimpack
vi O/OptimPack/build_tarballs.jl

    In this file, at least update version number and SHA256 code for the source tarball:

    sha256sum optimpack-${VERSION}.tar.gz

  3. Check that products can be built (from the Yggdrasil/O/OptimPack directory):

    julia-1.7 build_tarballs.jl

    You may upload the products (in the directory Yggdrasil/O/OptimPack/products) to the release page on GitHub.

  4. Commit changes, push the branch to GitHub:

    git checkin -m "Update for OptimPack version ${VERSION}" build_tarballs.jl
git push --all

  5. As suggested by this last command, create a pull request at https://github.com/emmt/Yggdrasil/pull/new/optimpack.

Prerequisites

Developers under Debian/Ubuntu need the following packages: build-essential, gcc or clang, autoconf, automake, autotools-dev and libtool.

Initial step size/scaling

The k-th iteration of a conjugate gradient method consists in updating the variables as:

x[k+1] = x[k] + alpha[k]*d[k]

where alpha[k] > 0 is the step length and d[k] is the search direction given by:

d[k] = -g[k] + beta[k]*d[k-1]

where g[k] is the gradient of the objective function at x[k]. Different formulae have been proposed to compute beta[k].

A line search method is used to find a suitable alpha along the search path. The initial value of alpha can be:

alpha[k] = alpha[k-1]*<d[k-1],g[k-1]>/<d[k],g[k]>

which is the rule used by Shanno & Phua (1978) in their CONMIN algorithm.

In quasi-Newton methods, the initial approximation of the inverse Hessian is often given by a simple scaling by a parameter gamma > 0. This yields to the following search direction:

d[k] = -gamma[k]*g[k]

Possible rules to guess the value gamma from the previous iterations are:

gamma[k] = <s[k-1],y[k-1]>/<y[k-1],y[k-1]>    (Oren & Spedicato)

alpha[k] = <s[k-1],s[k-1]>/<s[k-1],y[k-1]>    (Barzilai & Borwein)

where:

s[k] = x[k+1] - x[k]
y[k] = g[k+1] - g[k]

are the difference between successive variables and gradients respectively.