MW
iMW
iNEOS
PBB
Modeling/Condor
Condor
Globus
NEOS
iNEOS

An Interactive Environment for Nonlinear Optimization.
Marcel Good, Jean-Pierre Goux and Jorge Nocedal.
Collaborator : Victor Pereyra

The NEOS Server allows users to solve optimization problems remotely via the internet. Over the past few years, NEOS has proved to be very useful for hundreds of users from academia, industry, and government.

The current implementation of NEOS has, however, important limitations. It forces the user to submit all the data of his problem, and do so in rigid formats specific to each area of optimization. To solve nonlinear optimization problems, for example, the user has to submit Fortran or C files describing the objective function and constraints. This mode of operation prevents many potential users from accessing the servers, for the following reasons.

  • The data files or codes specifying the problem are often proprietary or confidential.
  • These files are too large (consider for example a 3D flow simulation).
  • The evaluation of the objective function and constraints is too time consuming to be delegated to a NEOS server.
  • The model is written in several languages or in a format that is not acceptable to the NEOS servers.
The iNEOS package is the first step in overcoming these difficulties. Its initial implementation is designed for solving unconstrained or bound constrained optimization problems. The task of evaluating the objective function (i.e. the simulation) remains in the hands of the user, and NEOS is only required to provide a new approximate solution. In nonlinear optimization, it is possible to generate an improved solution just by knowing the current iterate and the numerical values of the function and gradient at the current point. These arrays of real numbers do not reveal the nonlinear model to an observer of the data submissions.

Some of the nonlinear optimization codes in NEOS, such as L-BFGS, already have an internal client-server structure, that facilitates this mode of interaction. Here a driver computes the function and gradient values at a trial point and then calls the optimization code, which returns a better estimate of the solution.

iNEOS makes use of the following architecture to distribute the simulation (function and gradient evaluation) and optimization (generation of new trial point) on two different sites:

iNEOS has been implemented using CORBA. We demonstrated the effectiveness of iNEOS by performing an interactive identification of the properties of certain piezzoelectric materials.

More on iNEOS :
metaneos@mcs.anl.gov
Last modified: Mon Jul 3 23:12:21 CDT 2000