www.edn.comJune 10, 2004|edn69designfeatureBy Graham Prophet, EditorAt a glance ..............................70For more information............74 Despite considerable marketing hype a few years ago, fuzzy logicand neural networks failed to develop traction as selling points asfar as consumers were concerned. Yet, both still have their place in yourengineering tool kit. The two techniques areessentially unrelated, except that they bothprovide control methodologies to handlehighly nonlinear or poorly specified prob-lems, they both came to some prominenceat about the same time, and they both fad-ed from view in much the same way.Both neural networks and fuzzy logic as-pire to allow electronic systems, built withfamiliar circuit techniques or employingconventional computing technologies, toattack certain problems in a way that mim-ics human responses and abilities. One ofthe intimidating aspects of fuzzy logic isthe name itself, which has connotations ofimprecision. On the contrary, however,fuzzy logic is capable of precise responses.It allows systems built around Boolean log-ic, handling binary values, to work withimprecisely defined values that you mightexpress verbally as “more,” “less,” “high,”“low,” and so on. Although vague, thosevalues are amenable to intuitive humanunderstanding. (“Vague” is one of thosewords that has a special meaning in thespecial vocabulary.) You cannot definetheir meaning by using one parameter butwith reference to others; for example, if theproblem that you are exercising controlover is the separation between your car andthe car in front of it, the meaning you ex-press by “near” or “far” differs dependingon the speed at which you are traveling,whether you are accelerating or braking,and a number of other parameters.In implementing a fuzzy-logic system,you map values across the range of an in-put parameter, to membership of a set thathas some bearing on the control parame-ter you eventually want to calculate. Againusing the car example, you might decidethat for separation distance, 10m is nearand therefore a full member of the “prox-imity” value set and 100m is far and not amember of that set. So the input- valuerange 10 to 100m would correspond to amembership value of 1, reducing to 0, ofthe set. For simplicity, most hardware-based fuzzy systems use a linear mappingbetween the two limits, leading to triangu-lar- or trapezoidal-shaped “membershipfunctions,” the fundamental buildingblocks of fuzzy systems.You can then express logical relation-ships between the sets using verbal con-structions: For example, “If ‘proximity’ islow AND ‘speed’ is high, THEN ‘braking’is 100%.” Every application has multipleAxeon’s Vindaxchip implementsmachine learningon an array of 256parallel-processingelements.THOUGH NO LONGERHEADLINERS, FUZZYLOGIC AND NEURAL NET-WORKS ARE OPTIONS INTACKLING CHALLENGINGAPPLICATIONS.Fuzzy logic and neural nets:still viableafter allthese years?designfeatureFuzzy logic70 edn |June 10, 2004www.edn.comrules of a similar form describing the re-lationships between the controlled andcontrolling parameters. The process ofderiving an output from an input in fuzzylogic, “inference,” is not a calculation;rather, you convert real-world inputs tofuzzy values, identify the applicable rulesand derive a “weighted average,” andtranslate this information back to a real-world (“crisp,”the complement of fuzzy)value. In technological terms, the processof deriving outputs from a series of inputvalues with fuzzy logic involves “fuzzifi-cation,” followed by inference, followedby “defuzzification.” Figure 1 shows atypical fuzzy-system controller.Those who come to fuzzy logic from abackground in conventional logic aresurprised that you can obtain controlloops of high performance with a sim-plified design process. Long-time advo-cates of the technology acknowledge thatthe name “fuzzy” has not been helpful inthis respect. Fuzzy logic remains a simpleand quick method of producing controlsystems that can operate with few re-sources—especially in the case of non-linear systems, in which it is difficult towrite a precise model on which to base aconventional control loop. Using fuzzylogic, you can also interact directly withinput values, which consumers can de-scribe in “human-comprehensible”terms; you can label your washing-ma-chine buttons for laundry that is “lightlysoiled” to “very dirty,” for example.You can implement a fuzzy-logicscheme either through software develop-ment, leading to code that runs on a con-ventional microprocessor or microcon-troller, or you can use a microcontrollerthat has a hardware fuzzy-inference en-gine module. You can also source high-level mathematical-modeling softwarethat includes fuzzy-systems design.NEURAL NETSNeural networks, unlike fuzzy logic,seek to reproduce the versatility of thehuman brain in recognizing the end-to-end, input-to-output behavior of a sys-tem without understanding all theprocesses taking place within it. Taking asa fundamental model the interconnec-tions of nervous systems within thebrain—neurons and synapses—neuralnetworks have the attributes ofmemory and learning. In ap-plying a neural technique to a system,you show the network many examples ofknown-correct input/output-value pairs.In its learning mode, the network createsnetwork connections with weighted val-ues to match the data you provide andstores the values for the weighted con-nections that achieve the correct result.By exploring the whole input/output-value space, the network “learns” to pro-vide a correct response to any given in-put stimulus, without formally modelingthe processes comprising the original sys-tem. An essential trick in designing aneural-network architecture is to achieveconvergence; that is, as you show it suc-cessive input/output examples, it buildsthe ability to model the complete valuespace and does not “forget”the examplesit previously learned.Neural networks are therefore partic-ularly applicable to complex problemsinvolving difficult-to-measure or diffi-cult-to-model parameters—in particu-lar, processes that have a high degree ofnonlinearity. Although much discussedin academic circles, only limited exam-ples of pure neural-network control haveever appeared in volume-productionembedded-control applications, al-though many have done so as custom so-lutions to individual intractable controlproblems, in areas such as industrial-