TRON

From ASCEND

NLA
QRSlv
CMSlv
IPSlv
Opt
NLP
CONOPT
IPOPT
TRON
MINOS
NGSlv
OptSQP
DAE/ODE
IDA
LSODE
DOPRI5
LA
linsolqr
linsol
LP
MakeMPS
Logic
LRSlv

We are starting some work on building support for the TRON NLA solver. It seems unlikely that we will be able to distribute TRON in binary or source form, but it is available for free for 'internal research' purposes from http://www-unix.mcs.anl.gov/~more/tron/index.html

From the above page: TRON is a trust region Newton method for the solution of large bound-constrained optimization problems. TRON uses a gradient projection method to generate a Cauchy step, a preconditioned conjugate gradient method with an incomplete Cholesky factorization to generate a direction, and a projected search to compute the step. The use of projected searches, in particular, allows TRON to examine faces of the feasible set by generating a small number of minor iterates, even for problems with a large number of variables. As a result TRON is remarkably efficient at solving large bound-constrained optimization problems.

We have a build script (SConstruct in the directory models/johnpye/tron) that allows TRON to be compiled as a shared library, which we propose to use in further work on this. Tested with tron-1.2.tar.gz downloaded from above.

We have a skeleton of the TRON interface written in the file asc_tron.c also in the above directory.

At this stage, it became clear that the problem format supported by TRON, namely for single objective function plus numerous box constraints, might make optimisation subject to equality constraints rather difficult. It was not clear how to proceed, so this is currently stalled.