Hi together,
when I worked in Roy Sambles' Thin Films & Interfaces Group in Exeter in 1993 I used an existing Fortran program for multilayer Fresnel modelling and improved its slo-mo performance to the point where one could change the parameters of the layers and see an almost instant graphical response. This actually gave rise to the "Inverted SPR" discovery.
Now, since then I was wondering if it wasn't feasible to have the computer scan a huge parameter space on its own and essentially doing a "curve discussion" on its own: A new phenomenon or quality can be described in natural language terms and so can known patterns be described (how are minima, maxima distributed, what are known limits, etc.) as well. The latter would have to be told to the computer and then it would only start spitting out results if it found something remarkable, something outside the known patterns. With today's computing power such a generic software tool would have produced the ISPR instantly! Now I would like to pass on the suggestion: create or find a method to describe and detect patterns in modelled data. Then let a computer scan the areas you find interesting? What do you think? Or is it standard procedure nowadays to do this?
Cheers from Switzerland
Marc
The most sophisticated and elegant way to address this, I think, is using work by Steve Grossberg [much on this site; most of the rest at his BU page]. Steve's work on vision uses "Adaptive Resonance Theory" principles to discover qualities and categories well-suited to your problem.
The down side is the enormous processing load required [meaning a problem has to present complexity and value enough to employ ART].
The positive is that the system is inherently flexible and learns with as much granularity as required.
Much of Steve's code is in MATLAB; some of it is posted on his site.
Hope this helps.
Sam
Marc. If I understand what you ask correctly, you have some form of computer code, with many parameters that describe a huge, computationally expensive space within which to search, then you might consider using a design of experiments approach to guide you in the direction you are after. Though not often taught, in the last 40 years or so, there has been a marked increase in the application of DOE methods to computer simulations of various types. If you are interested here is a starting point: Santner, T. J., B. J. Williams, and W. I. Notz, 2003: The design and analysis of computer experiments. Springer-Verlag. I have also attached a link to the book.
Let me know if this helps.
Jeffrey
Hi Jeffrey, thanks a lot for your help! I'll have a look. Best regards - Marc
Hi Sam, I thought I had answered to you from my iPhone earlier, but looking at the answers here I suspect that didn't go through. Thanks very much for your answer, too! Best regards - Marc
Marc. Do please let me know what you think of the idea. I would appreciate that.
Jeffrey
I'd like to propose an alternatiive: Genetic programming. The advantage with GP is that it can scan a huge space very fast, and it's pretty generic as all you need to provide is a fitness function (describing what you find interesting). GP has been successfully used by industrials to make patented discoveries. Of course, the downside is that it's kinda random, so you can miss something interesting, or not find it at all (although there are "guided" exploration techniques to somewhat alleviate this issue).
Deep learning techniques have proved useful in doing prediction. A combination of NLP, clustering and prediction would address this problem successfully
Trade-offs:
Speed v texture ["depth" can be illusory; detecting interactions and subtle dynamics allows understanding of data flavor but takes time]
Stability v Adaptability [are criteria clear and stable? is it important to detect emergent or unstable characteristics?]
GAs are faster; standard stats are more easily applied [and better understood]; recurrent hierarchical networks are slower and require more intense design but offer granularity and flexibility and can be audited [to a limited extent]
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Computer Graphics and Animation