You fix the optimal solution? You expect the system model to stay the same
> over time and different inputs?
In my previous post "The same process can be repeated for other conditions (other points in input space) and other "analytical equation-solution" couples are memorized."
You are expecting a lot from an industrial
> unit. The grand idea of finding an overall equation for a unit is nifty,
> but quite often very oversimplified. Typically there are several competing
> phenomena, each of which contributes somehow to the overall operation. Not
> all of these phenomena are known in many cases. Some are not even > observable.
I checked this approach on the comprehensive model of a telecommunications network. It works pretty well for the 30 variables.
For complex systems, even those which look very simple to the untrained eye,
> it could fall rather flat.
To me, all live creatures on earth use this approach to be able to survive. We name it "a learned expirience". I only proposing a method how to find borders of multiple "good enough" solutions for different situations in the very fast way. BTW, what is a "good enough" is defined by a system designer. Creatures who select this definition in wrong way are not going to survive.
Setting one(?) optimal state ignores the vagaries of the situation. What looks
> optimal on one process could be screwing a downstream process up very badly.
Again, in my previous post: "The same process can be repeated for other conditions (other points in input space) and other "analytical equation-solution" couples are memorized."
How does your system differ in effect from dynamic matrix control? Is it
> as good as the dynamic matrix control with an online coefficient correction
> system? That watches the process and learns continuously from it, rather
> than fixing knowledge from one(?) test for all time.
In my previous post: "The second phase is an open-loop control itself. During this phase condition of an object (process) is monitored and current point into input space is defined. Corresponding to it "analytical equation-solution" is interpreted and solution is executed. It is still be fixed until object (process) input still being inside of the corresponding to this solution input area."
>
> > In 1995 I have published two articles in which I have proposed a new
> > approach to the Open-Loop Control. (1."Some problems with the design
> > of self-learning open-loop control systems." European Journal of
> > Operational Research. 1995. 2. "Input set decomposition and open-loop
> > control in telecommunications networks". 1995 American Control
> > Conference, Seattle, 1995).
> >
> > I think that any laboratory, that has a model of some object(process)
> > or can manipulate an object(process) itself and this
> > model(manipulation) allows performance optimization of an
> > object(process), can verify methodology I am proposing.
> >
> > In brief, the methodology consists of two phases.
> >
> > The first phase is learning. During this phase, for same set of object
> > (process) conditions (a point into a multidimensional input space)
> > performance of the controlled object (process) is opitimized and
> > optimal solution becomes fixed. Then, analytical equation for the
> > outer surface of expending area in the input space, in which found
> > solution is still be considered as "a good enough", is quickly
> > defined. (See "Input set decomposition and open-loop control in
> > telecommunications networks". 1995 American Control Conference,
> > Seattle, 1995). "Analytical equation-solution" couple is memorized.
> > The same process can be repeated for other conditions (other points in
> > input space) and other "analytical equation-solution" couples are > > memorized.
> >
> > The second phase is an open-loop control itself. During this phase
> > condition of an object (process) is monitored and current point into
> > input space is defined. Corresponding to it "analytical
> > equation-solution" is interpreted and solution is executed. It is
> > still be fixed until object (process) input still being inside of the
> > corresponding to this solution input area.
> >
> > Execution of both phases can be easily computerized.
> >
> >
> > How I can see it, this approach has a very broad application
> > potential:
> >
> >
> > 1. It will be possible to create self-learning procedures that provide
> > sufficient functionality for controllable objects, for example, robots
> > in unknown environment and, using continuously intersected
> > controllable areas, give those objects adaptive features.
> >
> > 2. It can be defined a sub-optimal trajectory of transferring a system
> > from one state to the other via continuously intersected controllable > > areas.
> >
> > 3. It will be possible to develop of a set of sub-optimal emergency > > plans.
> >
> > 4. Unstable systems (processes) can be controlled.
> >
> > 5. Diagnoses (for example, medical diagnoses) can be enhanced. > >
> > 6. Process of an image and voice recognition can be improved.
> >
> > 7. It will be possible to avoid some accidents by controlling human
> > activity and preventing controlled system from its drawing outside of
> > a multidimensional safety area.
> >
> > 8. It will be possible to recover from some disastrous situations by
> > using prepared in advance sub-optimal recovery procedures.
> >
> > One can continue this list of possible applications
> >
> >
> > I will be glad to provide assistance to anybody who will try to use
> > the proposed methodology.
> >
> >
> > Ziny Flikop
> >
> > snipped-for-privacy@earthlink.net.
> >
> >
> > P. S. 1/ Since "Input set decomposition and open-loop control in
> > telecommunications networks". 1995 American Control Conference,
> > Seattle, 1995 is short, I can fax it to anybody who will request it. > >
> > 2/ Please excuse my grammar, English is my second language.