BETTER THAN THE REAL THING
As you many know, I'm a member of the CIO ( Computer Integrated Operations ) committee of the National Center for Manufacturing Sciences ( NCMS ), Ann Arbor, Mich. Recently we had a CIO meeting at Kodak to decide the fate of nations . And I pontificated hugely about modeling and agents. Jim Heaton and Van Dyke Parunak beat me soundly about the head and shoulders because, as usual, I was locked onto the latest Morley romance, and would listen to little else.
Associates say I have only about 20 percent actual content in my talks. The problem is, they don't know which 20 percent it is. So they have to listen to the whole lecture. Even over the dinner table or on the ski lift.
Anyway, my talk centered around the several ways modeling can be done.
One way is to make assumptions in then examined for comparison with the real world; in other word, an "experimentally" developed model; in other words, an "experimentally" developed model. Controls engineers observe the actual process and tweak a model until the simulation behave the same as the real process. The model may be the starting point for model construction.
Often, changes introduced by engineers are motivated more by intuition, experience, and hunches than by academic "soundness". The proof of the pudding is the behavioral correspondence to the real systems, not the approach.
Examples of experiment-based model include the "Bullet" train in Japan. General Motors has a truck paint shop that is optimized for pull-though scheduling and process control using the approach. So-called generic algorithms embody the experimental approach since they make random changes to a model, observe their effect, and integrate successful changes into the model. Artificial intelligence and artificial life studies also use experiments in real time, as does life itself.
An older mechanism for modeling is the deductive method. This approach is common in research settings. It depends upon the accuracy and thoroughness of our theoretical understanding of the physics of the process. Theories are, at best, only a rough approximation, and a theory based model may work only for simple situations. A model based solely on theory may not in fact represent what actually happens in the process.
The third approach is the usual compromise. Inductively we can monitor the actual system with new math techniques. By reconstructing the geometry of the system, engineers have a model based on behavior, rather then theory. This approach has been applied extensively to formally chaotic systems, where theory is invariably inadequate to predict performance.
In general, model-based control incorporates a computational model of the process into the control system, rather than just reflecting the parameters and architecture of control. This approach appears to be more robust over a wide range of parameters. It does require that the engineer have an explicit model available.
Over the last several years, NCMS has funded Delphi studies for ascertaining the near future of technological thrusts for industry. Software is at the head of the class. All indications for the future involve modeling and/or agents. These studies herald an ongoing revolution in software. As with the hardware in the computer revolution, we can expect to see modeling in control systems change from an art form to a true engineering discipline supported by the pillars of scientific understanding.
Thanks to James Heaton and Dr. Van Dyke Parunak for their input in puttting together this column.
As appeared in Manufacturing Systems Magazine June 1995 Page 10
References - Table of Contents
|R. Morley Incorporated
614 Nashua Street, Suite 56
Milford, NH 03055-4992 USA
Tel: 603-878-4365 FAX: 603-878-4385