WHEN DETAILS MATTER: APPLICATION OF NEUROCONTROL TO STANDARD BENCHMARK PROBLEMS
Paul J.Werbos, Ludmila M.Dolmatova
In the last few years neurocontrol has made enormous progress, in terms of new engineering applications,
new mathematical designs and ideas, and new links to the brain. The success of real life applications depends
on a vast variety of factors, from general design ideas down to programming details. Sometimes it is not easy
to find the real reason for success or failure in a particular application, because the usual approach to neural
modeling is somewhere between art and science. So, it's time to demystify the process of design, from idea to
implementation to interpretation of results. There exists a set of benchmark problems described in [1] which
became a de facto standard for exploring the possibilities of new modeling and control methods. However,
most of these problems (even though they were intentionally designed for instructive purposes) can be
implemented in different ways, which complicates the comparison of results. The goal of this paper is to
demonstrate some "underwater rocks" in the application of neurocontrol methods using standard benchmark
problems as a commonly available touchstone.
[Contents]
|
