Dennis S.

Bernstein

“Mind the gap. Mind the gap.’’

II a recent survey 111,27 researchers provided inputs con- is not an accurate description oracademic researchers. Although
cerning the state of research i n systems and coiitrol and :illhut one of the survey respondents 11 J hold positions in acade-
were asked to give their opinions on the m j o r challenges mia, many have exlcnsive industrial and governincnt experi-
facing thc control community. Among thc challenges ence. Furtherinorc. several of the respondents in (I I are strongly
identified was the need to bridge the gap between theory in favor ol‘aggrcssive cfforts Lo bridge the gap. 1 am encouraged
and practice. by their views l o present some concrelc suggestions of my own.
In Lhis article I speculate on sonic reasons for the exi~tcnccof Pinally, the views I present arc intended to suggest how the
the gap and provide concrete suggestions for bridging it. Spe- academic side might contribute to bridging the gap hetwecn the-
cifically, I alii inlerestcd i n the Sollowing questions: ory and practice. However, tlicrc is much the industriiil side can
1. What is the evidence for the existence 0 1 lhc gap’? do as well to solve this problem. I helieve there is a correspond-
2. What is lhe extent n i t h e pip? ing burden on control proclitioners lo articulate their needs and
3. What is the significance d t h e gap Cor systems ;ind control provide guidance and feedback to the research communily.
research‘!
4. What factors have contrihuled to the gap’! Evidence for the Gap
5. What technical research problems are pertinent to bridging 1 will comment only hriecly on the lirst three questions. Al-
the gap’? though inany control rescarchers and practitioners would proha-
6. Why is il important to bridge the gap? bly agree that a Lheorylpracticc gap exists, the extent of the gap is
First, a few disclaimers are i n order. My perspective on these the subject o i inucli debate and call only be cstiinated by anec-
questions is froin acruspace engincering and reflects my experi- dolal evidence. Chwacterizing and quantifying this gap would
ences in academia and industry. Furthermore. this nrliclc is not require nontrivial eflort and is bcyond the scope or this article.
intended to he either a defense or a critique o l “academic” rc- Although there w e researchers who are quick to point out suc-
search in control technology for aerospace engineering or any cesses oE modern systems and control theory rescarch, 1person-
other branch of enginccring. ally believe that lhc gip 011 the whole is large and warrants
According to my dictioniiry, “acadeinic” incans “very serious introspecliiin by thc research community.
learned hut inexperienced i n the world of practical reality.” This The significance of the gap Lor systems and control research
is ii complex and sublle issue. Here I note that (I) basic research
has always benefited from the influence of applications, while
(2) in llic long run, the most important developments are those of

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:33:39 UTC from IEEE Xplore. Restrictions apply.
hesic research that have deep and long-lasting inllueoce, leading At tlic same timc, the opporluiiities affiirded hy conceptual
to the conclusion that (3) a bdance between theory and applica- a d v ~ ~ ~inconlrolasedifl‘icult
ccs t o g u ~ s pltcim
. bedifficultforrc-
lions is essential. Unfortunately, the timc constants for basic se- searchers to demonstrate and quantify thc advantages o l a new
search and useful applicalion are gciicrally quite different. In idea in control whese the dcsign proccss is complex and these ase
fact, new idcas can require a long time from conception to cx- tiumerous tr:idcoffs that inksact in intricate wiiys.
ploitation, while the time pressures of applications can divert at-
tcntion froin potentially valuablc long-term solutions. 2. W h a t a r e the risks ofncw control methods?
Next, I’ll discuss some factors that have contributed to thc In many ;ipplicetions. especially i n aerospace, thc control sys-
gap’s existence (queslioo 4). Thcn I’ll exaiiiinc various systems tem is critical to system operation. This is ii two-edged swosd. On
and control issues and their relevance to bridging the gap (ques- the onc lnmtl, control system technology is essential and therefore
tion 5 ) . Finally, 1’11 end by commcnling on thc importiuice of will command high priority when it is required. On the other hand,
closing the gap (question 6). i f the control systcm fails, the system may be lost, and thus appli-
cations in which human lives os grcat cost is at stake call for well-
Why Does the Gap Exist? tested methods over novcl tcchniques. Risk reinforces the inestia
To shed some light on the existence of tlie gap, it is wortli- assnciatctl with the acceptance of ncw contrul technology.
while to spcculale on some of the reasons for its existence. Nevertheless, there are many potential applicalions of coiitsol
technology that are not casily designed ;way. These include
1. Dn control citgineers need inndcrn sysiems a n d llight control with unusiial configurations (cantcd tail fins,
control thcory? oblique wings, tailless aircraft) and nontraditional control appli-
Generally, new technology is used in practice only when cations (compressor and comhustor control, active vibration
there isaclearcostorpeifiirmancebenefit.’lhcrc;irelwosides to control, high angle of attack flight). However, each of these q -
technology development, namcly, lechnology push and market plications entails risk and thus, (lcspitc potential cost and perfor-
pull, whcre “market” can rcfcs to either commercial or military/ inancc advantages, inust undergo cxtciisive development bcfore
government applications. In aerospace technology, the latter is it caii be transitioncd to practice.
usually the technology driver. Although necessity is often called
the mother of invention, the extent l o which necessity contrib- 3. How do publicatinns contribute to i h e gap?
utes to bringing about truly new developments is a murky ques- Bridging tlic gap, fsnm tlie researcher’s pcrspective, requircs
tion. (Some inventors believe that neccssity is actually tlic that new ideas becoinmuiiicatcd toengineers who are inaposilion
daughterof invciitioii [21,p. 125.)Despileclearnecds, new tech- lo apply them. I’lllbcus on publishing, whcre there are three main
nologies often require a gcncration (or more) to mature to the avenues, namely, journal papers, conference papers, and books.
poinl of practical uscfulness. The reality is surely a combination Much effort in acadctnia is devoted to publishing injournals,
of both push and pull. which are striclly limited to demonslfiibly new ideas. Consc-
I personally place the pointcr to the side of tcchnology pusb: qucntly, journal papers arc extremely terse and arc generally
Engineers oftcn invent the possiblc and look For opportunities writtcn Ens other researchers, not practitioners. Authors orjour-
they can develop and exploit. The most important and funda- nal papers arc rarely given much spacc to provide background
mental developments areofteninwardly motivatcd. Aiter all, the and sell-contained, pedagogical exposition that could render
Wright brothers were not funded by thc Air Force, and them was their papers inure readable by practitioners and nonspecialists.
no pressing need for miinncd flight. Nor was the transislor devel- This avemie is also slow, typically taking thrcc ycars from
oped out of necessity. In fact, its potenlial wasn’t even recog- submittal to puhlication
nized at firs1 in the U S . (although il was in Japan). As another Conference publicatinns are mose timcly hut are even inore
example, the mathematician G. I-Iardy took satisfaction in the tersc than journal papers (they o k n serve as summaries o l re-
uselessness of nuniber thcory, yet few technologies are hotter to- sults), and they arc not widely available to engineers who do not
day than data encryption. On the other hand, some of the most attend thc conrerence. Howcvcr, CD ROM proceedings are
important conceptual developmcnts have arisen from the desise much more accessiblc than bulky hasd-copy versions and should
to address real problems. Fourier analysis arose from heat flow allow larger distrihulion. Ipersonally hope that once paper pro-
problems for which Fourier was ridiculed, and thermodynamics ceedings are eliminated, authors can be allowed more publishing
arose from the boring of cannon [ 3 ] . Yet physicists are often spacc to develop iicw idcas.
called on to contribute to revolutionary techiiologies such as ra- Books provide an opportunity for researchers to expand and
dar and nuclcar weapons. illustrate new concepts at great length, which is not possible i n
In control engineering practice, “need” is a nebulous concept. journals. Few books written by academic rcscarchcrs, however,
Rarely do engineers consciously design and build a system that consider sufficient engineering detail to be directly usable to
truly cannot bc controlled with existing concepls. In many appli- praclicing engineers. Some “how to”contso1 books are available
cations, the plant can be redesigned and additional seiisors and on spccialized topics such as motor servos, but fcw of them truly
actuators can be implemented to rcnder the control problem advancetlieslaleoTtheast incontrolpractice(noris it theirintent
more manageable. Fully actuatedmechanicirl systems are agood to do so).
exampleoftbis kindofdesign. Other brute force solutions can hc
devised as well. Forexamplc, pcrfosmance specifications can be 4. How can publications b c rendered m n r c usable?
lowered, and hardwarc and manpower budgets can be increased A control enginccr considering ii new control algorithm lor a
(or, more commonly, projects can be canceled), thus eliminating potential iipplication needs to know the answers to the following
opportunities for new and innovative, and tliercfore risky, ideas. questions:

Detiemhsr 1999 65

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:33:39 UTC from IEEE Xplore. Restrictions apply.
 1. What problcms does the control algorilhm address (stahili- gineering praclice. Control architecture dcsign rcfcrs to the se-
zation, dislurhance rejection, tracking, ctc.)? lcction of sensors and actuators that need lo be specified to
2. ’To what class of plants does the algorithm apply? achieve a control objective. The dcsign of the control architec-
3. What modeling information is needed to design and tunc ture and associated hardware usually depends nn a solid under-
the controller? standing of the relevant physics along with detailed analytical
4. What is the structure o l the control algorithm? modcling. In fact, analytical inodeling at this stage in the control
5 . How is the controller tuned? engineering process is extrcmcly cost-cffcctive since it reduces
6. What sensorlprocessorlactuator hardwarc is nccdcd to im- the need to fahricate and test multiple prototypcs.
plement the controller? On the other hand, control architecture design is often only
A satisfactory conlrol paper will provide answers to all these loosely coupled with coiitrollcr implcmentation, that is, the
qucstions in a clear and accessible manner. If this inforination is choice of the coiitrol algorithm and its tuning (parameter sct-
buried in the paper, the reader may have to expend considerable tings). In fact, modeling for controller implcmcntation usually
cffort and thus may become discouraged rrom further consider- requires information that is distinct from the information needed
ing the method. For example, if lhe control law is given in trans- for control architecture design in both type and detail. Forcxam-
formed variables, t h e potential uscr must unravcl t h c ple, although finite-element modeling and compuiational fluid
traiisforniatioiis to determine the roles of the various parameters. dynamics may provide important information fin sensor and ac-
Although this information may be a detail to the writer, it is vital lnator design and placemcnt, thcse modeling techniques usually
to thc practitioner. cannol provide lhc type of detail nccdcd for controller imple-
As a furlher example, a practitioner may he interested in a mentation, such as plant phase at crossover.
special case of a general procedure. The reader may not have The distinction between modeling lor control architecture de-
time to rccognizc the applicability of thc proccdurc or to undcr- sign and modcling Cor controllcr implcmentation clarifies thc
stand how the result specializes to useful, special cases. 1person- role of distributed parameter models in control design. Such
ally likc the “closcd-loop” or “ s a n d w i c h model of paper models provide a starting point for the furnicr but havc littlc relc-
writing: concrete, abstract, concrete. That is, motivate the paper vancc for the latter.
with a concrcte or specific problem, work out thc thcory in an ab- None of tliesc remarks are intended to minimize the impor-
slract or gcncral coiilexl, and, rinally, relurn to the concrete or tanceolcithcr analytical or data-based modeling in control cngi-
spccific problem. In general, a paper is most useful to a control neering. In fact, holh kinds ormodeling arc cxtrclncly important,
practitioner whcn it providcs and dcmonstrstcs operational pro- and they arc thc responsibility of thc control enginccr. Howcvcr,
cedures for implementing the methods under conditions that arc it is important to recognize what modeling information is needed
explicitly stated and knowable at each stage of the control engineering process.
The following remarks are intended to emphasizc the rcle- At early stagcs in control architccturc dcsign, thc modeling may
yBncc of various topics and issues that I believe have signifi- h e largely analytical and hypothetical, wlrerci~scontroller imple-
cance Tor bridging the gap hetween theory and practicc. Many of mentation must be strongly linkcd to thc hchavior of a spccific
these topics and issues have been extensively studied by the sys- hiirdwarc rcalization.
tems and control community, while others have not. My ohjcc-
live is to emphasize those aspects that may have some bearing on 7. Rediice the dependence of analytical and data-hased
the gap. modeling for controller implemcntation.
As discusscd above, analytical modeling is essential and
Bridging the Gap: Modeling Issues valuable for control architecture design, hut il has serious short-
comings for controllcr implcmcntation. Although control archi-
5. Don’t trivialize stabilization. lecture design often consumes the bulk of coiitrol engineering
Although it sccms trivial to say so, unstable plants arc much effort, advances i n control algorithms can rednce the need for
more difficult to control than stable plants. Yet unstable plants both analytical and data-based modeling for controller imple-
are often viewed simply as linear plants with oiic or more open menlalion.
right-half-planc polcs. (A plant with a chain o l integrators or The ob,jective of robust control is to achieve perlormalice for
imaginary poles is also onstable, hut less seriously,) a givcn lcvcl of modcling unccrtainty; huwcvcr, robust control
1 helieve the distinctioii between stable and unstable plants is Fails to rednce the dependence on eilher analytical or data-based
vastly imdercmphasized in the research literature. An unstable modcling significantly. Although robust control methods do not
plant provides almost n o opportonity for on-line identification require a precise model of the nominal plant dynamics, they do
and thcrcrorc niust rely heavily on analytical modcling and cx- require that all uncertainty be quantified, and thc construction
trapolation from stahle regimes. Unstable plants are unforgiving andverificationofsnch adetailed uncertaintymodel may require
in the sense [hat once large deviations occur, saturation limits substantial analysis of prcdiction modcls and test data.
inay prevent recovery. Furthermore, linearizing a nonlinear UII- The main drawback of robust conlrol is that il treats uncer-
stable plant may obscure the actual saturation rccovcry limits, tainty as a static qnantity, which forces the control cnginecr to
which areinvariahly smaller than Lhoseofthe linearized model. sacrificc pcrformancc for robustness. Ultimatcly, robust control
requires that the controller gains be decreased to account for un-
6. Distingiiish between modeling for control architecture certainty, thus reducing performance. The inability of a robust
design and mndeling for controller implementation. conlroller lo learn makes this tradeoffunavoidahle.
Control architecture design and controller tuning are strongly Finally, control engineering must accept the possibility that
interrelated, but they are effectivcly distinct t;uks i n control en- any givcn p h o t can change in an unexpected and unpredictable

66 IEEE Control Systems

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:33:39 UTC from IEEE Xplore. Restrictions apply.
manner during operation. In fact, seemingly small physical Bridging the Gap: Control Issues
changes can have a large effect on plant response. For exeiiiple,
as luhrication dissipates, bolts loosen, mass distribution 9. Respect the distinction between continuous time and
changes, and components wear, the plant dynamics inay changc discrete timc.
significantly. Thesc unprediclable changes are the responsihility A cruel fact ofcontrol engineering life i s that m o s l of the sys-
of the control engineer. lndccd, a major reason Cor implemcnling tems we need to conlrol operate in coiltinunus time whilc the
a feedback control system i s to achieve performance i n the pres- controllers we implement on digital computcrs operatc in dis-
ence of nnccrtainty, and not all unccrtainty caii be characteri7.ed crcle time. It can be dirficult to reconcile the continuous and the
or predicted. discrete; sometiincs lhey behave like oil and waler. Conse-
qncntly, the inlerfacc helwecn contiimous- and discrete-time
8. Exploit identification for controller implementation. systcms i s a lricky business, and i t can liave a significant cffect
No matter how well analytical modeling can he pcrSorined, on control syslein performance. Tlrc literature abounds with
some identification i s always needed. Real hardware abounds transformations betwecn coiititiuous-time and discrete-timc dy-
with unmodelablc effects and high sensitivities. In addilion, namics such a s lustin's, exponcntinl, and hilinear. However,
modeling a systcm iiipieceiiiealfarhion i s of limited use for con- lhcse are increly convciiienccs that don't address stahility and
lrollcr implemcnl&m. since components can interact dynanii- pcrformancc in a reliable manncr.
cally in complicated ways due to spurious fecdhack paths and The continuous/discreteiousI(liscretegip is hcrc to stay, sincc lherc i s no
unexpectcd interactions. The need Cor idcntification and hard- revolution in analog controller technology oil the horison. Fur-
ware testing i s crucial, and end-to-end identification i s desirahle llierniore, even i l we could implcmcnl continuous-time coiitrol-
whenever possible. Obviously, identification i s only meaningful leis, our identification incthods opcrale on discrete-timc data to
after the systcm has been construcled and data are obtained. produce discrete-time models. Identification in continuoils lime
The ability to perform identification depends on the natiirc cif is not a scrious prospcct. So willi analytical modeling and classi-
the plant as well as on thc environment. Idcntificalion uf the un- cal control in conliimous time and with idcnlilication in discrete
controlled plant i s generally not feasible ifthe plant i s opcn-loop timc, it's n o surprisc that thc control literaturc often appears
unstable. Tn that case, a stabilizing controller i s needed, which schizophrenic.
inay require analylical modeling or adaptive methods. In addi- It i s lernpting to believe lhat for sul'ficicntly fast computers,
tion, the presence of ambient dislurbances can liinil lhc ability to discrete-timc systems am be treated as continuous-time systcms.
identify and adapt. Io this case, idcntilication accuracy may be Howcvcr, there arc fundamental distinctions bctween discretc-
low and the results o f the idcntif'ication may bc nonrepcalablc.IS time and conlinuous-lime systems. For example, discrctc-timc
ambient disturbances can be climinated, idcntirication i s much control has an inhcrenl hindwidth limitation iiiiposcd by the
casier. (Engines can be turned off,whereas turbulent wind noise sample rate. A delay in continuous timc is an irrational exponen-
around a flying aircraft caiinot.) Identificalion and adaptive sta- tial function, whereas in discrctc time i t i s riitinnal (one nice hen-
bilizalion in the prescnce of exogenous disturbances presents a c l i i of discrete-time models). In addilioii, a nilpotent linear
severc challenge to control engineering. For this problcm, the discrctc-limc syslem has finite scttling lime behavior, whereas 21
control engineer i s forced to rely more heavily on analytical linear conlinuous-time syslem cannot settlc in linitc time. ( A
modeling. lime-invariant conlinuous-lime syslein that seltles in Sinite time
Numerous issues of theoretical and practical significance are necessarily has non-l.ipschitzian dynamics IS].) Finally, the bc-
associated with identification. Since identification i s difficull in havior o f the systcm betwccn sample instants can affect closed-
the prescnce of fast and slow dynamics, ii delta-operatoridentifi- loop pcrlhmancc. Il'the sample interval i s short, the intersample
cation theory would bc useEul 141. Choosing good identificalion behavior should be henign Whether this effect can be ignored in
signals, especially i n the presence of ambient disturbances, i s a practice i s an open question.
problcm of practical interest. I n some sciise a good identilicetion There arc Suiidaincnlal obstacles in sampled-data control lliat
signal i s "far" from a disturbance signal. Coding ideas may be must he treated carefully. First o f all, sampling and rcconstruc-
useful in this regard. tion devices, which provide thc interfacc between the continu-
Nonlinear identification i s largely an open area o f research ous-time and discrete-time worlds, have time-varying dynamics
with considerable practical importance. Since a l l rcal systems with inherent limitations. Arbitrarily fast sampling i s an unrea-
are nonlinear, i t i s overly simplistic to apply linear idenlil'icaiion sonahle expectation since faster hardware merely encourages
methods and expect that any such method will produce a man- cnginccrs lo consider ever iaster planls or morc coiiiputalionally
ingl'ul linear model. I suspcct that difficulties observed with lin- intensive conlrol algorithms. Furthermore, Cast sampling can
ear identification methods are due lo unniodeled nonlincarities cause numcrical prohlems with poles aggregating near I,The
as muchas sensor anddisturbancenoise. We must also recognize dclta operator provides a praclical solulion to this problcm [4].
and admit the possibility o f systems that have nonrcpeatablc be- Similarly, zcro-ordcr-hold signal recon~tructionis a timc-vary-
haviorduc lo sensitivedepcndence on initial conditions, ambient ing operation lhat produccs spurious harmonics. Suppressing
disturbances, and complex dyoainics. thcse effects i s oftcn required lhrough addition;il Ciltering.
Finally, slatistics has been underutilir,ed by (he conlrol com- Aliasing i s aproblcm that arises duc to sampling. Folding of
munity a s a whole for analyzing identification and pcrformancc signals and noise i s an unavoidable cfSect of aliasing, and i t i s
data, although statistical malysis h i s been sccing increasing in- rarely nccounled for explicitly in control theories. Aliasing d s o
terest i n processconlrol. The analysis olany data withoutcarcful causes phasc shifts iit lowcr frequencies that caii destahilim a
statistical analysis i s naive at bcst. system It is important in practice to dcterininc sample rates and

Ilecember 1999 67

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:33:39 UTC from IEEE Xplore. Restrictions apply.
design anti-alinsing filters with niiiiiinal phase lag to supprcss A control cngirieer who has invested in control systcm herd-
thcsc effects. ware i s nftcn inlercslcd in achieving tlie bcst possible perfor-
Nonlinear systcms are difCicull l o trmslate clcanly into dis- inaiice from (he chosen h;irdwarc. Whether or not fuel or cnergy
crete time. For example, finite escape time can occur in continu- constraints arc presenl, this goal may require that the acluittors
ous time, but i t has no direct countcrparl in discrcte time. opcratc at or near saturation Icvels. Hcncc saluratiun limits arc
Capturing nonlincar physics in discrctc time i s a iiontri\,ial chal- wily reginns to he avoided, but rather may bc songht
Icnge, especially sincc our training and iiitiiitioii are based in so as to iiiiixiiiiizc use o f llic available control input.
continunus time. Exacl di etization of some coiitiiiiious-tiiiie The dislinclion between stable and unstable systems i s im-
Inodcls i s discussed in 161. portanl when addrcssing saturation issues. If the plan1 is open-
loop slahlc, saturation i s only an issuc when performance is
10. Distinguish roal-timecomputing from off-linecomputing. quantified, siticc the zero crmtrol i s unsaturxled and stabilizing.
Many of the rcspondcnts polled in thc survey [I I discussed
On the othcr Iiand, glr,hal st;rbiliz;ition of plants with right-half-
pliiiic poles i s impossible in lhc prcsence of saturation. Thcrc-
the ondcrutilization and potenlial cxploitation o f rcccnt ad-
vances in computational power. One application of powerful Core, inaxiinising thc domain of iittriiction i s the primary objec-
live Tor tilistable plants. Sincc the domain nf attraction i s
computers is to solve very largc order Lyapunov and Riccxti
ieccssarily boundcd, a rare disturbance or high magnitude can
equations. This i s reminiscciit ofthc "big drum"approach: Prim-
perlurh llic state and rendcr the equilibriuin unrecoverable. (Big
itivc tribes wishing to communicate with llic uutside world
ships.) This problem i s critical when consid-
might conceivcof ever largcrdrums. These arc examples ulmis-
ilback on imsl;tblc syslems.
guitlcd technology scaling.
rcndcr linearization misleading for unstable
The usefulness of any computation must bc cvalualed in liglil
pliints. Specifically, linearizing a nonlinear iinstablc plant may
of the accuracy of the underlyiiig data. Most nicasiircnicnls arc
obscure the aclual satoralion recovery limits, which arc invari-
good to only ahnut 0. I W , nnd mxny paramelers (not to mcntion
ably smaller than those of lhc linearized model.
physical cffccts) are significanlly more uiiccrtiiiii. (A stein werii-
ing on the unrcliability of tlatii i s given in Chapter 27 of 171.)
12. Reeogniac limitations due to sensor noise.
M,dssive
..' computation based 011 erroneous data or hypolhclical
I t i s important to stress that a11 real signals are corrupted by
models may havc qualitativc valuc for insight, but the actilal
noisc, and this noisc limits tlic ;ichievabic performance. Noise
numbers produced will likely havc little cnnncctioii with reality.
caii arise Irom the sensurs and all associatcd electronics, and its
(Ifinsight i s the goal, this i s not n problem; it'siust inipnrtant l o
e rarely known prior to implementing thc cnn-
keep thesc goals distinct.) Computing with unccrlain data hiis
trol system hardware. 111particular, lhc noise may be due to de-
been largely a neglected tupic in the scientific community, al-
l a i l s o f grounding and shielding, whose efkcts are difficult to
lhough tlie robust control community (to its crcdit) has given it
predict bciitrc the plant and controllcr have been constructed.
serious allention.
I%eding back conlrol signills feeds back tlie sensor nuke as
In control engineering, large-scale computing i s relevant Ibr well. Therefore, i C thc disturbance i s narrow band but its spcc-
plant tilid control architecturedesign, which i s largely a qualita-
trum i s nnt kiiown in advence, a control engineer might be in-
tive and hypothetical process. Such computing i s pcrforined olT clined to u s e a controller with broadband gain. Howcvcr,
line and uftcii occurs before the plant and control system arc con- feedhack in a frequency range where the plant disturbance i s no1
structed. This computing i s usually performed fur the sake (if prcsent will amplify sensor noisc. Thus, lhc presence o l sensor
modeling, hut i t i s suspect as a viable approach to controller im- noise forces tlic control engineer to limil control gains and bend-
plementation. On the other hand, controller implementation caii width. This design issue i s o k n ignored in control design theu-
he enhanced by the capability Tor rea-time, on-line computing. ries since the sensor noise spectrum i s rarely known from
ldentification and performance assessment during control sys- analytical modeling. Purthermorc, in LQG theory, (his con-
tem operation for adaplivc control i s one o f the main hencfici;i- straintisdifficult tolian(llcbecausenarrow-band noiscgivesrise
ries of significant tcal-time computing ciipability. tu ii singular estimation problcm. 111any cveiit, real noisc i s
There i s no real tradeofrhetwecn on-line and off-line comput- surely more iiisidious than idealized noise mudcls.
ing. They are distinct tasks that use different kinds of informa-
tion for different purposes. Off-line computing i s based oii static 13. Emphasize the distilletion hetween smooth and
and usually limiled inlormation about the system, wlicr~ason- nonsmnoth nonlincarities.
line computation has continual access tu data from the truc sys- Although linearity over a range i s an oxymoron. i t i s never-
tcm and its infinitely rich physics as it behaves in possibly unprc- theless useCu1. However, nonlinear effccts assume greatcr im-
dictable ways. portancc a s perlormance requirements become morc stringent.
Many control methods considcr smooth nonlinearities. which
11. Always recognize saturation. are lincarizahle ncarcquilihriaand haveanincrcasing effect over
Often thc first nonlinearity encountcrcd by tlie control engi- a larger raiigc o f operalion. Geometric nonlinearities in robotics
nccr is saturation. (Here I am rcferring to amplitude saturation. are thc prototypical cxeinples olsuch nonlinearitics. Control the-
The second nonlinearity encountcred i s rate satoration.) 11 i s a mists tend lo think of lhcsc nnnlinearitics a s being well known
universal nonlinearity that will ncver be circuinvciited by any and iimeiiable tu global translormation techniqucs.
tcchnological developmcnt. Saturation i s a lincarizable O I -I On the other hand, many control applications arc of a prcci-
linearity that has a global impact on tlic system but has little cf- sion iialurc where lhe objective i s to produce highly accnrale mo-
fect on the local behavior of the systcm. tionover small amplitudes. I n this regime, the nonlincarities tend

6X IIZEB Control S y s t e m

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:33:39 UTC from IEEE Xplore. Restrictions apply.
to be nonsinooth (thal is, not linearizablc) and possibly discon- 16. Tlicre are no details in control engineering.
tinuous. Friction i s ii cnininoii cx;imple of ii nmsinooth nrin- Tlicrc arc no "details" iii control engineering, since cvcn the
linearity. In additiun, nonsinooth iioiiliiicarilies inay possess inmt insigniliceiit "detail" niay prove l o be important. All cngi-
memory as well, forcxamplc, sliction and backlash or hysteresis. nccring ultimately hinges on dctails, because real systems intist
Hysteretic inonliiicaritics are usually semistable subsystem [ R I he built from rcal, iiiiperfcct (no1 mathematic;il) components and
with multiple equilibria where subsystem convcrgcnce i s fast intist opcrale under real (nonideal) conditions. This point holds
relalive to the remaining system dyniimics. The incmory charac- all the inure kir cuntrol enginccring, a complex technology that
tcrislic i s inerely the lrajectury-dependent set o f equilibria thal depends on many intcrrclatcd aspects. Thc smallcsl "details"
lhe subsystem converges l o during quasi-slatic upcralioii. Thc such 21s noise, quantiziition. drift, bias, crosslalk, roundi~ff,
phase lag nature of such nonliiiearitics renders theiii polentially aliasing, nonlinearities (local or global, smooth o r ~ion~mouth,
destabilizing eveii iil low sigiial :impliludes. mcmorylcss or noninemorylcss). saluration (amplitude or rate),
Cliissical control theory discusses both smooth and non- delays (known or unknown, fixedor variable), inotlcl errors, scii-
smooth nonlincarities with iiii emphasis on the lormcr through sorlactualor dynamics, state cuiislraints, and system changes can
absolute stability theory. However, nonsinootli nonlinearities adversely tifScct conlrol system operation. Tlic gap belween thc-
seem l o be iiiorc prcvalenl in applictitions. In Pact, while largc- ory and praclice can be narrowed by systems and conlrol theory
iimplilude motions can often bc slowed down without major loss that recognizes the iinportancc o f lhese issues (and surely inany
of performance (and this i s oflcn done i n practice), lack ulpreci- others) ancl addrcsscs them in a meaningful and usefill way.
sion in small-ainplitudc applications can serioosly degrade the
v~iluc of the system. In other words, large robols can bc opcrated 17. Reduce the dependence on modeling.
more slowly if ncccssary (although it may iiot be desirablc to do
I bclicve that one cif the inain culprits i n the theorylpraclice
this), but alack nf precision in it inwhining task illay not he tolcr-
gap i s the modcling dependence imposed by many control mclh-
able at any speed o f operation.
o d s . 111fact, the modeling requircmenls iniposcd by model-based
In general, nun~mootlinonlinciiri~iesarc casier to idcnlify be-
control mctliods constitiitc a severe iinpcdimcnl to the applica-
cause the amplilude range i s smaller. However, thcsc non-
bility uf modern control thcury. Analytical modeling i s esscntial
linearities come in a wide variety of types, they iiiay bc hidden,
for control architcclure design, but it intist he used with care for
and thcy inay changc draslically and uncxpectedly over differcnt
Ihe purposc OS controller implemeiitation.
operating ranges. On the other hand, smooth nunlinearitics arc
dil'licult to identify because of the riingc of opcration rcquired to The success o f PlD tuning inclhods relative to moderncontrol
collcct data. Control theorists lend to view siicli nonlinearilies as mcthods is ii remindcr of this dependence, while inodel predic-
well known becauseof thc analytical nature ofclassical nicchaii- tive control bascd on identificatioii i s cqually successful for the
ics. I n applicalions such as [light control over a largc envelope, same reason. Therelore, i t scciiis that tlie extcnt to which a con-
identification of global nonliiiearitics can he extremely difficult. trol inethod i s used in practice i s proportional to its iiirxleling re-
quircmcnts, making this issiiea key factor i n llieexistencc ofthe
gap. Tlic first slep toward remedying this prublein i s to distin-
Final Observations guish between modeling for control architecture design beforc
syslem construction and iiiodcling lor controller implcmentation
14. Kemembcr the transients.
(usually thrnogh identification) after system construction.
Control theorists lhave afixation with cquilibrioin-related bc-
havior. Wc seek the steady stale because it i s easy to chariicleri%e
and provides a safe Ihaven. Lyapunov stabilily lhenry, whiclicoii- 18. W h y bridge the gap?
tinucs to provide arich hiinling ground [XI, 191, lids spoilediis. In I have left this qiieslion for last because i t i s the inusl funda-
engineering practice, il i s often Ihc transients that matter. Colli- incnlal and most dillicult. Thus iar, Ihave suggested that tlicrc
sion avuidaiice i s a good motivating cxample. But dcaling with arc pragmatic rcasons for closing the gip between theory and
transienls i s not easy. As S . Ulainonccsaid, "The infinite wccan practice. Thc triinsilion of new idcas and techniqucs Lo applica-
do immcdiatcly; tlic finite takes a little longer." Lions ultimately justifies the cost of basic rcsc:irch Although
tlicre certainly have bccn succcsscs in tlie application of inoclern
15. Feedback entails risk. idcas lo technology, the penetration of modern idcas in many ap-
Most engineering disciplines are open loop i n tlic sense that plications seems to he fairly limited. Serious attcntion to techno-
errors are iiot amplified. A 20% error in the strength of a strtic- logical nccds and constriiinls i s esseiitial for understanding and
tiifill membcr rcmiiins j u s t hat, and a 50% inargin will cumpco- correcting this state of affairs.
sate for the error quite nicely. (The Hoover Dam was designed Ontlieotlicrhand,itisiinportanttokeepinmindthatb a ,'m re-
wilh a3x safety faclor fortlioscofyou livingdownstreain. Asfiir search i s iiicanl to be high risk in terms ofp;iyofl'. Control cngi-
as 1 know, it i s not stahilized by Sccdback loops.) 111contrast, iiccrs with prujcct deadlines rarely have the luxury o f pursuing
feedback affccts dynamic behavior, and " s i i i ~ i l l " errors ciin pro- uncoiiveiitiuiiiil idcas with uncertain return. 1h;it i s the d e ofac-
duce arbitrarily largc iindesirahle cffccts (such iis inslahility). aileinia, whcrc rcscarchers can instantaneously shin research di-
Attempts Lo guard against this scnsilivity assume that modeling rcctioiis mil pursue new idcas without management approval
unccrteinly i s known, yct lhe cuntrol system n i i i s t havc tlie abil- (unlike the usual case i n induslry), or pursue ii novel idea for
ity to copc with unexpcctcd changes iis well. Sincc control sys- years until i t i s sufficiently dcvcloped to havc tcchnological
tenis are often critiail l o operalion with significenl losses in thc v~iluc.What i s largely lacking i n the academic setling i s mutiva-
event o f failure, tlie ability to cnpe with unexpected changes i s tion from reiil applicalions. I believe that exposure to such moti-
the responsibility of the control engineer. vation, even lo a limited cxtcnl, can havc a significant impact on

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:33:39 UTC from IEEE Xplore. Restrictions apply.
 Author's Note
I'dliketosnytliatiny Ihinkingonthisissur
was inlluenced by my cxperience i n industry
(Harris Corporation, Aerospace Systems Di-
vision, 1984-9I ) aiid academia (University of
Michigan, Aerospace Engineering Depart-
iiiciit, 1991-present). But I think it would be
more accurate lo say thiit thc greatcst ioflu-
ciiccs on my carcer were ( I ) the Erector Set 1
got when Iwas seveii years old and (2) [lie tiinc I spent fixing my
car wlicn I was in college.
Although tinkering can be valuable, what aiiiales iiie most is
the ability of the human intcllccl to solve rea-world problerns
IIir(iugh abstraction, and yet I see that much of what cnginccrs do
is ruotcd in cmpiricisni. L think that a lo1ol'enginecring is actu-
ally an art, hut today's art oRen bccoincs tomorrow's science.
And what is surprising about this knowledge is its comprcssibil-
ity. An insight os br~ikthr(iughthat t ~ i ak lifetime to achieve
inay heconie common knowlcdgc fur the next generation. The
greatest iiitelleclual achievcmcnt of all limc (the alphabet) is
kniglit to prcschoolcrs, while the second-grealcst ([he calculus)
is kiught to high school studciits.
My latest [avorite qoole is by Thomas Hdison: "Wc don't
know a millionth of one percent about anything." I think this is
worlh keeping in iiiiiitl at tlic closc of the inilleiiiiium.

70 IlWR Control Systems

Authorized licensed use limited to: MIT Libraries. Downloaded on January 14,2021 at 01:33:39 UTC from IEEE Xplore. Restrictions apply.