•  
• EDN Europe MAGAZINE
•  
• Free SUBSCRIPTION
• Subscription Renewal
• View Media Kit

• EDN
• EDN Australia
• EDN Asia
• EDN China
• EDN Japan

EDN Europe's Editor Graham Prophet posts a selection of comments and insights prompted by the many items of industry news and rumour that cross the editorial desk or are gathered on his frequent travels to interviews, press conferences and events around Europe - and further afield - and somehow never find their way to the magazine or the web site, recovering some of the information otherwise lost in the noise level...

Wednesday, April 01, 2009

Fuzzy thinking

| Permalink | Email This | Comments (0) |

| Digg This | Slashdot This | add to Del.icio.us |

One of the unpredictable phenomena that I’ve noticed from time to time is that a particular subject will disappear for a long time, then it will pop up more than once in a short time. In exactly this way, I’ve had two readers contact me in the last few days asking about fuzzy logic and neural networks. Prior to that, I’d heard little mention of either for quite some time.
I wasn’t really able to assist either reader very much. As is well-known, editors (some of them, at least) have heavily cached processors (I’m referring to their internal organic processors) – you only stand a chance of getting sensible answer if you ask them about something they happen to have been linvestigating in the recent past. Ask about a subject that has been flushed from the mental workspace, and all you get is a cache-miss signal.
I did recall that the only bit of neural hardware I had seen recently was that being developed by Recognetics – which company, when I went to look at its website, turned out to have become Cognimem. Most users of neural/fuzzy techniques seem to run them in simulation on conventional machines, and probably start out in packages such as Mathematica (which handles fuzzy operators) or Matlab (that has a fuzzy toolbox).
Lots of academic work has looked at the possibility of hosting neural algorithms on FPGAs: the attraction of mounting adaptive logic structures on re-configurable hardware is fairly obvious. I conclude it’s probably not all that easy, however, because you don’t very often see it done in practice.
For a sector of the subject that has seen, perhaps, a higher ratio of research to deployment than some other areas have, when my attention is drawn to it anew, a couple of thoughts arise. The first is, why have these techniques not made more of an impact in practice, than you might have expected given their apparent versatility? I think part of the reason may be that they occupy a rather small niche. Both – neural and fuzzy – have applicability, in somewhat different ways, to systems that are governed by some underlying order: but in which that order is either too complex, or has too many variables, to be easily and fully characterised. Most real-world control problems, meanwhile, are amenable to a conventional approach: you can fairly readily establish the rules that govern their behaviour, and the computation required to carry out control by a conventional program is not too great. The subset of real-world problems with enough order to be addressable via a neural/fuzzy solution, but too little to use a conventional control loop, may be too small to justify commercial development.
Does that hypothesis hold water?
The other question that arises when I get more than one such query in a short time is, am I missing something? Is there something stirring out there in neural/fuzzy territory that I haven’t noticed? Please, let me know……

Related entries in: Processors/ Software/ Tools |

Post a comment

* Display Name
* Comments
* Anti Spam security code	To pervent spamming, please enter the anti-spam code as shown above to proceed with the submission