Wickens and Kessel 1980

8/6/2019 Wickens and Kessel 1980

1/14

Journal of Experimental Psychology:Human Perception and Performance1980, Vol. 6, No. 3, 564-577

Processing Resource Demands of Failure Detectionin Dynam ic SystemsChristopher D. Wickens and Colin KesselUniversity of Illinois at Urbana-Champaign

The information-processing channels, proprioceptive versus v isual, that are usedto detect changes in the response of dynamic systems are investigated using aloading-task methodology. Conditions are compared in which sub jects eithercontrol the dynam ic system (M A mode) or monitor an autopilot controlling thesame system (A U mo de). Failure detection in these tw o modes of participationis evaluated w hen sub jects perform the task alone an d concurrently with either atracking loading task or a me ntal arithmetic-mem ory loading task. T he formertask disrupted MA detection but not AU detection, whereas the converse re-sults were obtained with the mental-arithmetic task. The results, interpretedwithin the framework of a structure-specific resource theory of human attention,suggest that AU detection relies exclusiv ely on processing resources associatedwith perceptual/central-processing stages. M A detection in contrast relies onseparate-processing resources residing in a response-related reservoir.Dynamic systems may be characterizedas receiving comm and or disturbance inputsfrom environmental sources and generatingoutputsmathematically describable system

statesthat are deterministically related tothose inputs w hen filtered through a systemtransfer function. This description character-izes a wide variety of systems with whichhuman operators interact, ranging from th eextreme complexity of the modern nuclearpow er plant (R asm ussen, 1979) to the relativesimplicity of the bicycle (VanLunteren, Note1), or garden hose. Relatively sophisticatedtechniques are available for describing thedynamics of such systems in control the-ory (Toates, 1975) or state space (Rouse &Gopher, 1977) representation. Basic research,howev er, has not dealt extensively with th ehuman operator's conception, understan d-This research was sponsored in part by Air ForceOffice of Scientific Research Grant 77-3380 (AlfredFregly was program manager) and in part by NationalScience Foundation Grant BNS 78-07860 (JosephYoung was program manager). A portion of this re-

search represents a part of the PhD dissertation ofColin Kessel, currently employed by the Israeli AirForce.Requests fo r reprints should b e sent to ChristopherD. Wickens, Department of Psychology, Univers i tyof Illinois, C ham paign, I llinois 61820.

ing, or internal model of the dynamic sys-tem (Jagacinski, Burke, & M iller , 1977; Jag-acinski & M iller, 1978; Pew, 1974; Sheridan,1970; Smallwood, 1967.) The experimentalevidence that does exist however suggeststhat the human operator's internal modelof the system often represents an idealizedabstraction, employing various heuristics,rather than a faithful representation of thesystem itself.An impo rtant property of real-world sys-tems is the mean s by which supervising in-dividuals maintain a level of system familiar-ity. In a monitoring (autopilot or AU ) mode,control and regulation of the system is ex-erted by an automatic control mechanism,and the human operator's exclusive role issupervisory; a visual display of the systemstate is monitored, and this information isused to update an internal model of its char-acteristics. In this capacity the operator mustof course be prepared to intervene manu allywith appropriate control actions in the faceof any abnormality or discrete change in sys-tem response.

In contrast to the AU mode, in a manualcontrol (M A) mode, the hum an operator isactively engaged in regulation and control ofthe system throu gh continuous m anual inter-action. The consequences of control inputsCopyright 1980 by the American Psychological Association 0096-1523/80/0603-0564S00.75

564


2/14

PRO CESSI NG RESO URCE DEMA NDS 565can b e observed on status displays appropri-ately filtered by system dynamics, whereasthe control inputs are perceived via proprio-ceptive channels. In this mode of course,an additional function is now required be-yond the supervisory/monitoring of systemstatus, this be ing the control and regulationof system output in response to imposedenv ironmental criteria (e.g, m aintaining air-craft position along a glide slope or processplant production at a specified outpu t qualityand rate).Systems naturally vary considerably in theextent to which AU versus MA participationis involved. Those of extreme complexity,such as the nuclear reactor, function prima-rily in the AU mode, whereas simpler dy-namic systems such as the automobile orsingle-engine aircraft require primarily MAparticipation. Residing in an intermediatecategory are systems such as the modern je tair carrier in which alternate mode s of pilot(MA) or autopilot (AU) control are avail-ab le. It is this kind of system that the presentresearch will address.Wickens and K essel (1979) compared theability of subjects to maintain currency witha dynamic system under supervision in thetwo participatory modes. The system was ofweighted first- and second-order dynamicssuch that system output function of t ime,O(t), was related to the input, i(t), by thetransfer function O(t) = a ] J i(t)dt + (1 -a) / i(t)dt. The fidelity of the sub jects' inter-na l model was operationally defined by thelatency and accuracy with which subjectscould detect changes in the system dynam-ics. These changes were implemented bystep increases in the value of a, increasesthat caused the system to behave in a moresluggish, less responsive man ner.1 It shouldbe noted that this particular transition didnot generate a discrete deterministic changein the system display (e.g., a step change inposition) but rather a more subtle alterationof the relation between input and output.

Two conditions were contrasted. In theMA mode, the operator controlled the sys-tem with a two-axis joystick to make itsoutput track a slowly moving target on thetwo-dimensional display. In addition to thiscontrol input from the operator, the system

was also perturbed by a Gaussian noise in-put, analogous to the buffeting effects ofwind gusts on an automobile or aircraft. Theeffect of this noise is to induce errors into thetask that must continuously be nullified. Inthe AU mode, the human controller's func-tion was replaced by an autopilot controllerwhose dynamic transfer function generatinga control input to the system in response tothe perceived error was of analogous mathe-matical form to that of the human operator(McRuer & Jex, 1967). In both conditionsthe operator observed the system responseand the commanded input on a cathode-raytube (CRT) display and attempted to detectthe transitions when they occurred.A major conclusion of this investigationwas that detection is superior (both moreaccurate and of shorter latency) in the MAas opposed to the AU participatory mode. Anumber of fine-grained analyses were per-formed on the data in an effort to identify thespecific information channels utilized in up-dating the internal m odel and thereb y in pro-viding transition-related evidence. In theAU mode, naturally, the visual modality(error position, velocity, and acceleration)could provide the only sources of informa-tion. However, in the MA mode, proprio-ceptive information was also available con-cerning the control inputs immediately de-livered to the system. Knowledge of theseinputs could potentially supply useful in-formation concerning system response to agiven manual input or supply information in-directly in terms of the change in controlcharacteristics required to track the modi-fied dynamics. The availability of this addi-tional channel, it was hypothesized, mighthave been responsible for the observedMA superiority.Information relevant to the cue utiliza-tion strategies employed was extracted fromthree kinds of analyses: (a) analysis of en-semble measures of display and control vari-ables sampled continuously over time for hitand missed transitions, (b) linear multipleregression of response latency onto thesevariables at discrete time points following

1 This change might correspond to the sudden lossof stability augmentation in a modern jet aircraft.


3/14

566 C HRIST OPHER D. WI CK ENS A ND C O L I N K E S S E Lthe transition, and (c) analysis of cumulativeresponse accuracy as a function of responselatency. This latter analysis follow ed the la-ten cy operating characteristic or cumulativeaccuracy function (C AF) concept introducedby Lappin and Disch (1972) and Wicklegren(1977), which assumes an underlying variablecriterion model fo r transition detection (Ear-ing, 1977; Gai & Curry, 1976). This modelpostulates that evidence from informationsources (visual and/or proprioceptive) con-cerning system state is integrated over timeand compared with a steady-state model; ifthis new information is assessed to be suffi-ciently discrepant from the steady-state in -ternal model (exceeds a criterion), then adetection decision is triggered. If the cri-terion is lax, the decision will be rapid butless accurate. Trial-to-trial v ariability of thecriterion should thereby reflect the timecourse of growth of information concerningsystem state.In their analyses of the C AF , W ickensan d Kessel (1979) observed a distinct dis-continu ity in the M A, as opposed to the AUdetection functions, suggesting th at th e for-mer were generated by two temporally dis-tinct processes. Using convergent evidencefrom th e other microanalyses, W ickens andKessel postulated these two processes tobe, respectively, the rapid accumulation ofproprioceptive information immediately sub-sequent to the transition and the slower in-tegration of perceptual (display) inform ationthereafter. The second (v isual) portion ofthe CAF appeared to be equivalent in boththe MA and AU functions.The purpose of the present ex perim ent isto provide confirming evidence for this dis-tinction between the utilization of visual andproprioceptive informa tion channels by ap-plying a dual-task methodolgy. The aspectof this methodo logy exploited in the currentapproach is the multidimensionality or ap-parent structural specificity of human infor-mation-processing resources. A n u m b e r ofinvestigators have argued that processingresources or attention does not residewithin a single undifferentiated reservoir b utis better described by a number of structure-specific reservoirs (e.g., Isreal, W ickens,Chesney, & Donchin, 1980; Isreal, W icken s,

& Donchin, 1979; Kantow itz & K nigh t, 1976;Kinsbourne & Hicks, 1978; Nav on & Gopher,1979; Roe diger, Knight, & Kantowitz, 1977;Sand ers, 1979; W ick en s, 1980). Tasks rely-ing on comm on reservoirs will interfere to agreater extent than those using separate res-ervoirs. Evidence for the structural comp o-sition of these reservoirs has been summa-rized elsewhere, and structural dimensionshave been described b y processing modali-ties, cerebral hemispheres, and processingstages (W icken s, 1980). From the v iew po intof th e present research, th e structural di-mension of resources of greatest importanceis that defined b y processing stages. Consid-erable evidence may be cited that suggestsprocesses involved in response organizationand execu tion, and those related to percep-tual encoding and me mo ry compete with eachother (between categories) for resources toa degree that is less than th e competitionfor resources of two processes within eachcategory (e.g. , Isreal, C hesne y, W ickens, &Donchin, 1980; Isreal, W ickens, & Donchin,1979; Kantowitz & Kn ight, 1976; Na v on &Go pher, 1980; W icken s & Kessel, 1979).In the present ex perimental paradigm, tran-sition detection in each mode is performedalone, and, concurrently with each of twoloading tasks. O ne is design ed to place great-est demands on perceptual/memorial pro-cesses and the oth er, on response functions.If detection in the two m odes relies on struc-turally different information sources, thenan interaction is predicted in which dual-task decrements in the MAmode, dependenton proprioceptive channels, should be great-est with the response-loading task, w hereasdecrements in the A U mode should be moreattributable to memory loading.

MethodApparatus

The basic experimental equipment included a 7.5cm x 10 cm Hewlett Packard Model 1300 C RT dis-play, a spring-centered, dual-axis tracking hand control(with an index-finger trigger) operated with the domi-nant hand and a spring-centered fingercontrol operatedwith the other hand. A Raytheon 70 4 16-bit digital com-puter with 24,000 character positions of memory andanalog to digital, digital to analog interface was used togenerate input to the tracking display and to process


4/14

PROCESSING RESOURCE DEMANDS 567responses of the subjects. The subject was seated on achair with tw o armrests, one for the tracking handcontroller and one for the side-task finger controller.The subjects' eyes were approximately 112 cm from theC RT display. The overall display subtended 3 ofvisual angle.Tasks

Pursuit-tracking task. The primary pursuit-trackingtask required the subject to match the position of acursor with that of a target, which followed a semi-predictable two-dimensional path across th e display.The target's path was determined by the summation oftwo no nharm onically related sinuso ids (.05 and .08 Hz)along each axis with a phase offset between th e axes.This produced a target that moved along the path of aslowly rotating figure eight. The position of the follow-ing cursor was controlled jointly by the subject's con-trol response and by a band-limited forcing functionwith a cutoff frequency of .32 Hz for both axes (seeFigure 1) . Thus , the two inputs to the system werewell differentiated in terms of predictability, bandwidth,and locus of effect (target vs. cursor). The control dy-namics of tracking task were of the form Oft) = K[(l -a) j i(t)dt + a j j i(t)dt] for each axis, where a is thevariable parameter used to introduce changes in thesystem dynam ics, K is the gain of the stick, and i is theposition of the subject's control input. The transitions,or simulated failures, were introduced by step changesin the acceleration constant a from a normal value of.3, a mixed v elocity an d acceleration system with a highweighting on the velocity component , to a = .9, a sys-tem that approximates pure second-order dynamics andrequires the operator to perform considerable differen-tiation and prediction to maintain stable tracking per-formance (McRuer & Jex, 1967).In the MA condition, the subjects manipulated thecontrol stick to minimize tracking error. In the AUcondition, the subject's role in the control loop was re-placed by an autopilot whose control dynamics simu-lated those of the human controller of the first-orderplant: a pure gain, time delay, an d remnant (low am -plitude Gaussian noise added to the control output)(McRuer & Jex, 1967; W ick ens , 1976). In a series ofpretests, the parameters of the autopilot were adjustedso that the root mean square (RMS) error of autopilottracking was equivalent to that produced b y the hum anoperator in the MA condition.In both the MA and AU modes, subjects were in-structed to detect failures b y depressing the trigger asrapidly as possible. If a failure was detected, the dy-namics were imm ediately reset to the pretransition level.If the failure was not detected after 6 sec, the dynamicsreturned to the prefailure level of a by a smooth 4-secramp. A discrete step return to normal was avoided inthis case to guard against the possibility that subjectsmight detect the return as a failure.Critical tracking loading task. In this task (Jex,1967), subjects were required to manipulate a spring-loaded finger control in the left-right direction withtheir left hand to stabilize a system with unstable,

SUBJECT CONTROL

TARGET PATH

Figure 1. Subject display showing primary-task an dcritical-task tracking and schematic logic of these tasks.

positive feedback dynamics. System output wasindicated b y a cursor presented in the middle of the maintracking display, and the control was to be manipulatedin such a w ay as to keep this cursor on a reference pointin the center of the screen. The difficulty of the criticaltask was manipulated by setting the value of theinstability constant X at values of .5 and 1.0. The highervalue of X produces greater instab ility, requires m orecontinuous control, and has been validated to demandgreater amounts of the operator 's limited processingresources (Gilson, Burke, & Jagacinski, 1978; Jex,1967). Critical-task performance was asessed by anRMS error measure.Memory-loading task. Over stereo headphonessubjects heard a sequence of prerecorded two-digitnumbers occurring every 2 sec. At unpredictableintervals (on the average of every 15th number), a tonewas presented and the subject was required to respondby subtracting the numb er seven from a preceding digitand verbally report th e answer. In the easy condition,seven was to be subtracted from the digit just prior tothe probe tone. In the difficult condition seven wassubtracted from the Digit 2 positions p rior to the prob e.The two levels of the memory-loading task thereforehad the following characteristics: Both required fewresponses b ut placed a continuous demand on memory.Performance was assessed as the accuracy (percentcorrect) of responses.


5/14

568 CHRISTOPHER D. W I C K E N S A ND COLIN KESSELProcedure

In the experimental conditions, all trials were of 2 minduration. To avoid predictability, a given trial mightcontain four, five, or six failures, an d these were pre-sented at random intervals within each trial, subject tothe follow ing constraints: (a) No failure could occurwithin the first 15 sec (subjects were so informed), (b) Aminimum 5-sec interval was imposed between the re-turn to baseline (following either a detection or a miss)and a subsequent failure, (c ) Task logic insured thatchanges would only b e introduced when system errorwas below a criterion v alue. In the absence of this latterprecaution, changes would sometimes introduce ob -vious jumps in cursor position.Both the critical task and the m ental-arithm etic taskswere defined as loading tasks. T hat is, instr uctio ns pre-sented to the subjects stressed that these tasks shouldb e performed as well under dual-task conditions (con-currently with failure detection) as under the controlconditions in which performance on the loading taskalone was assessed. Furthermore, it was emphasizedthat subjects should perform as well on the difficultlevels of these tasks as on the easy levels.A system of contingent bonuses, in which good failure-detection performance was rewarded only to the extentthat these loading-task performance criteria were met,was imposed to reinforce the effect of the verbal in -structions. This bonus system rewarded subjects fordetected failures (hits) and penalized them for falsealarms. Because the side task was designated a loadingtask, a 500 bonus was awarded if its performance wasmaintained at a constant level across al l experimentalconditions. The detection performance bonuses werealso contingent on side-task performance constancy.Experimental Design

Three variables were factorially crossed in a mix edmode between and within-subject design. Participatorymode (AU vs. MA) and loading task (critical trackingvs. mental arithmetic) w ere between-sub jects variables,and task load (detection only task vs. easy task vs.difficult-loading task) was varied within subjects. Allsubjects participated for 3 days. Day 1 was dev otedexclusively to extensive practice on both the loadingand detection tasks. Special emphasis was placed onfamiliarizing subjects with th e detection task, allowingthem to track (MA) or observe (AU) the dynamics inboth the failed an d unfailed mode an d providing themwith several identified examples of the precise nature ofthe failure. (See Wickens & Kessel, 1979; Kessel &W icken s, Note 2, for greater detail on the train-ing methodology.)During each of the two subsequent (experimental)sessions, sub jects received an initial sequence of 4 wa rm-up trials (1 without and 3 with the loading task) followedby 15 experimental trials. These consisted of 5replica-tions each of the three within-subject loading condi-tions. Only the data from these latter sessions are re-ported below.

SubjectsSix right-handed male students at the University ofIllinois were assigned to each of the four between-subjects conditions. All subjects had normal vision andwere paid at a base rate of $2.50/hr in addition to anyperformance bonuses that they might earn.

ResultsTo estimate failure-detection accuracy, theprobability of a hit, />(H), was computed asthe proportion of the 50 failures experiencedby each subject across the 2 days that werecorrectly detected within the 6-sec interval

following the occurrence of the failures. Es-timation of false alarm rate was more com-plex, given the undefined nature of the re-sponse interval. A variant of the method offree response (Watson & Nichols, 1976) wasthereby employed in which the total time pro-viding opportunities for false alarms (triallength minus the number of seconds of6-sec hit intervals) was partitioned into 6-sec false alarm epochs. The false alarm ratewas thereby computed to be the number offalse alarms divided by this numb er of falsealarm epochs.The hit and false alarm measures werethen combined to generate an overall mea-sure of detection efficiency. Because of areluctance to assume the parametric form ofthe underlying signal an d noise distribution,the nonparametric measure of the areaunder the receiver operating characteristic(ROC) curve [A(ROQ] was selected to in-dex detection performance (Craig, 1979;Green & Swets, 1965). This measure isderivable from a single point in the ROCspace by the formula:

A(ROC) = - P(FA)]and is also tabulated in McNichol (1972),from which th e present values were taken.Craig (1979) asserts that the A(ROC) mea-sure meets tw o important criteria for a sen-sitivity measure: It preserves the order ofperformance about which categorical judg-ments of superiority can be made (Norman,1964), and it remains independent of re-


6/14

PROCESSING R E S O U R C E DEMANDS 569sponse bias.2 Values of A(ROC) and themean response latency fo r each conditionare shown in Table 1.From the results of prior research (W ickens& Kessel, 1979) and as confirmed inTable 1, it is apparent tha t experim ental vari-ables influence both the latency and accuracyof the response. To capture the joint effectof both of these variables in an overall esti-mate of detection efficiency, the latency andaccuracy measures w ere linearly combinedfor each subject in a detection efficiency indexof the form/5 = lO A (R O C) - Latency(sec),an index that generates higher values as la-

tency is shorter or as accuracy increases.This efficiency index is justified b y con-sidering the failure-detection paradigmwithin the fram ew ork of statistical decisiontheory as Gai and C urry (1976) have don e. Itis assumed here that subjects aggregateevidence over time concerning the dis-crepancy between the sampled-system b e-havior and the internal m odel of a nonfailedsystem, until this evidence exceeds an in-ternal decision criterion. Detection ef-ficiency is reflected in the rate of aggrega-tion of internal evidence, independent ofthe criterion setting. Efficient detection willbe fast and accurate, and efficiency there-fore should b e reflected by an index withopposite weightings on latency and ac-curacy . The relative 10 to 1 weightings onthe accuracy an d latency measures are ofcourse somewhat arbitrary. In the aboveindex, they are determined by the relativevariability of the two measures observedby Wickens and Kessel (1979) and repli-cated b y Kessel and Wickens (Note 2).In this sense the two variables correspondroughly to standardized Z scores.Mean v alues of the detection-efficiency in-dex and of primary-task RM S tracking errorin the 12 experimental conditions are shownin Table 2. The table naturally does not in-clude tracking error in the A U mode, sincethese values were com puter determined andtherefore unaffected by the loading tasks.How ever, it should be noted that the meanRM S error in all AU conditions was equal to. 11, a value equivalent to that in b oth single-task MA conditions.

Table 1Mean Failure Detection Accuracyand Latency Values

Single task Dual taskTask Measure- D i f f i -Loading task mode ment Easy cultDetection accuracy (A(ROC)) a

Critical MA mode .90 .86 .86tracking task AU mode .84 .82 .83Mental MA mode .92 .92 .90arithmetic AU mode .85 .81 .81Detection latency (sec)

Critical MA mode 2.32 2.71 2.78tracking task AU mode 3.61 3.53 3.60Mental MA mode 2.39 2.30 2.29arithmetic A U mode 3.12 3.40 3.13Note. M A = manual. AU = autopilot. RO C = re-ceiver operating characteristic.a 1.00 = perfect accuracy, .5 = chance.

Tw o separate one-w ay analyses of variance( A N O V A S ) were performed on the primary-task tracking error data, one for each load-ing task. With th e critical-loading task , a re-liable main effect of task load was obtained,F(2, 20) = 46.0, p < .01. Tuk ey tests of thedifference between single- and dual-taskRM S error and between the easy and difficultlevels of the critical task revealed both ef-fects also to be statistically reliable (p < -01and p < .05, respectively). The A N O V Aperformed on the tracking data for themental-arithmetic group substantiated thatthe effect of this loading task on trackingperformance was minimal and not sta-tistically reliable (p > .10).Performance on the two loading taskswhen performed concurrently with the de-tection task is shown in Table 3. Single-task loading task performance was mea-sured only on the initial (practice) day andso is not included here. A nalyses of variance

2 Even if the A(ROC) measure is affected by bias,there appears to be little likelihood that the differencesin this measure across conditions might reflect biasrather than sensitivity shifts. Mean false alarm ratewas relatively constant with a mean of .097 and stand-ard deviation of only .028 across all conditions.


7/14

570 CHR I S TO PHE R D . W I C K E N S A N D C O L I N KESSELTable 2Primary Task Failure Detection andTracking Performance

DualLoading task Mod e Single Easy Difficult

Detection efficiency index (1 0 accuracy-latency)Critical M A 6.70 6.03 5.85tracking task AU 4.71 4.68 4.72Mental M A 6.81 6.82 6.73arithmetic AU 5.45 4.82 4.93

Primary tracking RM S error (Proportion of scale)Critical

tracking task MA .11 .16 .17Mentalarithmetic MA .12 .11 .11Note. MA = manual. AU = autopilot. RMS = rootmean square.of the loading task data indicated reliablemain effects of task difficulty fo r both tasks:critical task F(l, 10) = 32.18, p < .01;mental-arithmetic task, F(l, 10) = 19.36,p < .01. In neither A N O V A did the effectsof mode or the Mode x Load interactionreach the .05 leve l of statistical reliab ility.The particular experimental hypothesis un-der investigation addressed the differentialimpact of the two loading tasks on detec-tion in the two modalities. The assumptionwas made that the two tasks would placedemands on structurally different resourcepools, and exa mination of the m anual track-ing performance suggests initially that thiswas the case. Th us, the critical trackin g load-ing task, which has been argued elsewhereto depe nd on response-related resources(Isreal, C hesne y, W ickens, & Donchin, 1980;Navon & Gopher, 1980; Wickens & Kessel,1979), disrupts primary tracking perform-ance. The latter is not however affected b ythe mem ory-loading arithmetic task. Y et,th e memory task is not without resourcedemands (e.g., it is not automated), as canb e seen by its disruptive effect on AU de-tection (Table 2). This point will be ad-dressed later.In assessing the impact of the loading taskson detection performance, it is essential toconsider separately the effects of the addi-tion of the loading task (difference between

single- and dual-task performance) and ofthe manipulation of loading-task difficulty.Following a discussion of this distinction b yRoediger et al., (1977), it must be empha-sized that increasing loading-task difficultycan only b e assumed to demand a greaterquantity of processing resources (from eithercommon or separate pools from the prima rytask) if loading-task performance is main-tained at a constant level. If this conditionis not met, as in the data of Table 3, thenit is impossible to assess whether more re -sources were in fact allocated to the moredifficult vers ion of the loading task .Clearly if primary-task performance falls atth e higher loading-task difficulty level, thisassumption is safe. However, if primary-task performance remains constant, as isgenerally apparent in Table 2, then the effectof task difficulty remains ambiguous. Eitherseparate resource pools are involved, or com-mon resources are used, but resource allo-cation to b oth tasks remains uncha nged, andonly the performance on the manipulatedtask varies. Since this ambiguity is presentin the current results, given the noncon-stancy of loading-task performan ce, the ef-fects of the difficulty manipulation will notbe considered further.Fortunately, similar problems do not ariseconcerning the effect of the introdu ction of aloading task. Performance of this task atwhatever level must, b y definition, consumemore resources than nonperformance.3

3 Roediger, Knight, and Kantowitz (1977) havecautioned against interpreting th e decrement fromsingle- to dual- task performance as a reflection ofresource demands because of other qualitative dif-ferences between these conditions that may be con-founded (e.g., changes in strategy, motivation, orstructural interference). The position taken here isthat strategy differences may in fact be in effect, b utthese should b e constant across al l loading tasks an dtherefore not influence a comparison of decrementsthat will form th e basis of the current analysis. Moti-vational confounds should b e just as prominent withincreases in task difficulty as with concurrent taskintroduction and again can be assumed to be roughlyequivalent across loading tasks. Finally, w e argue thatth e distinction between structural interference andcapacity interference is incorporated within the frame-work of structure-specific capacity theory (W ickens,1980). It is jus t as meaningful to consider interferenceeffects as tasks are introduced and to impose discrete


8/14

P R O C E S S I N G R E S O U R C E D E M A N D S 571Therefore, in the curren t results, data fromthe two dual-task conditions are pooled, andmean values of latency, A(ROC) , and thedetection-efficiency index were computedfor each subject across the two dual-taskconditions. These data were then subjectedto three-way repeated measures A N O V A Swith loading tas k , task load (single v s. dual),and participatory mode (AU vs . MA)serving as the variables of interest. The jointeffects of the three indep ende nt variables onth e detection-efficiency index are portrayedin Figure 2.The analyses of all three dependent vari-ables yielded a qualitatively similar patternof results, reinforcing the visual pattern evi-dent in Figure 2. In each A N O V A , only threeeffects reached or approached statistical re-liability. These were the main effects of par-ticipatory m ode: latency, F( l, 20) = 26.17,p < .001; accuracy, F(l, 20) = 19.0, p


9/14

572 CHR ISTOPHER D . W IC KENS AND CO LIN KESSELLoading Task Effects

The present results seemingly support theassertion that different information channelsare employed in the two detection modali-ties and that these rely on functionallyseparate processing reservoirs. Competi-tion by the loading task for resources inthe appropriate reservoir will selectivelyinterfere with detection performance, assuggested by Figure 2.Since the three-way interaction of Figure2 is critical for this interpretation, and sincethis interaction is dependent on the con-stancy of performance in the AU mode withintroduction of the critical-tracking task an din the MA mode with introduction of themental-arithmetic task, it is essential to con-sider if other explanations, not related toresource pool separation, could be offeredto explain these constancies. In particular,Roediger et al. (1977) have suggested twoalternative explanations for equivalence ofsingle- and dual-task performance. Theseare related to automation and resourceallocation:Automation. If one or both tasks areautomated or are in a data-limited region ofperformance, such that withdrawal ofprocessing resources leaves them un-changed (Norman & Bobrow, 1975), thenlittle interference can be expected with theirconcurrence even if common resources areinvolved. In the present data, neither ex-planation appears likely, since both tasks(M A and AU detection) were shown tosuffer performance decrements when pairedin different combinations (Figure 2) . Suchdecrements by definition provide evidencethat neither task is data limited.Resource allocation. It is possible in thetw o conditions under scrutiny (A U withcritical task, MA with mental-arithmetictask), that subjects maintained a constantsupply of resources from an undifferentiatedpool to the detection task both in the pres-ence and absence of the loading task. Thu s,the loading task would be the one to absorbthe decrement imposed by concurrence.This seems unlikely. I t would suggest thatoperators were not performing the detectiontask w ith ma ximum effort und er single-task

conditions. Yet, the nature of the bonussystem imposed made it financially re-warding for them to do so. Furthermore,this interpretation would require that theloading task was performed less well in thepresence of the detection task than in itsabsence. This also did not appear to be thecase. Single-task data collected for themental-arithmetic task showed a mean per-centage correct response of 87%. Whenperformed concurrently with MA detection,this percentage actually increased to 89%.Although single-task critical-task data werenot available (as noted, they were onlycollected on the first, practice day), it isimportant to consider that critical-task RMSerror in the AU detection condition wasreduced to 3.5% of scale. This translatesto only 5.4 minutes of v isual angle and doesindicate exceedingly accurate perform ance,certainly not much worse than could beexpected under single-task performance ofthe critical task.Given that both alternative interpreta-tions of the absence of effects in Figure 2can seem ingly b e ruled out, the explanationattributable to separate resource poolsremains viable. This explanation was ofcourse substantiated further b y the absenceof effect of the mental-arithmetic task ontracking. It is instructive also to considerthe pattern of results that would haveemerged had either or both of the assump-tions underlying the separate resource poolinterpretation (that resources are undiffer-entiated or that identical information chan-nels are employed in both modes) beenfalse. In the former case, the predictedordering of dual-task decrements would berelated strictly to the difficulty (resourcedemands) of the tasks, in which case oneloading task and/or one modality shouldconsistently show a smaller decrement thanthe other. In the latter case, each modeshould be affected similarly by the twoloading tasks. Neither of these patterns ofresults were observed.One final artifactual explanation thatneeds to be considered is whether the re-duced M A detection performance with th ecritical tracking task might have been at-tributable indirectly to the greater primary


10/14

P ROCESSING RESOURCE DEMANDS 573tracking error in this condition. This ex-planation would postulate that the noisiertracking display would mask the failuresignals to a greater extent and therebyrender them more difficult to detect. How-ever, consideration of the nature of thefailures provides some evidence against thisinterpretation. Since the failure itself ismanifest as a differentiated system response(greater acceleration) to a displayed errorsignal, the larger this error is allowed tobe at the instance of failure, the more salientthis differential response will become.Thus, although the particular artifact cannotbe ruled out altogether, since the con-founding of display error with loading taskin the MA condition is an inevitable con-sequence of the tasks performed, the natureof the failure suggests that it may not bea major source of effect.M A Superiority

In addition to the three-way interactiondiscussed above, the other notable aspectof the results concerned the main effect ofprocessing mode on detection performance,a finding that replicated at least three pre-vious investigations (e.g., Wickens &Kessel, 1979; Young, 1969; Kessel &Wickens, Note 2, Experiment 1). Thesource of this effect is of considerabletheoretical and practical interest in its ownright and so warrants some discussion.Furthermore, since the interpretation of thethree-way interaction is based on the sup-position that different channels are em-ployed in MA and AU detection, evidenceshould be provided to support the assertionthat the observed MA superiority resultsfrom the presence of the additional proprio-ceptive input.Wickens and Kessel (1979) have arguedfor the role of proprioception based on theirmicroanalysis of the detection and trackingdata, and the present results support thisproposition by the dual-task interferenceeffects. Yet, there exist other possiblecauses for MA superiority as well, threeof which shall be considered later.Despite efforts in the current paradigm toequate tracking performance of the human

and computer (by incorporating a computerautopilot that simulated the linear and non-linear elements of the human controllerand by adjusting these elements to equatetracking error), it is possible that the visualdisplays were inherently different. Thisvisual difference, rather than the presenceor absence of the proprioceptive channel,might have lead to the difference in de-tection efficiency. However, tw o lines ofevidence collectively suggest the visualequivalence of the two displays.Kessel and Wickens (Note 2, Experiment1) , using equivalent conditions to thosecompared here, assessed the mean velocityof the cursor just prior to failure occurrence(e.g., during normal dynamics). This mea-sure, which should reflect the total powerexerted on the control by the autopilot orsubject (a dimension of display characteris-tics independent from RM S error), wasessentially equivalent between the twoconditions. Further evidence that MAsuperiority still obtains even when displaysare in fact identical was provided b y Young(1969), who compared AU and MA detec-tion when th e display viewed by the AUmonitors was yoked to the MA display.That is, the autopilot whose tracking wasviewed by the AU detector was actuallyanother subject participating in the MAcondition.A second potential source of MA su-periority pertains to differential experience.This superiority might be a consequence ofthe greater familiarity with the systemdynamics, resulting from the active trackingexperience of the MA subjects, an experi-ence precluded the AU group. This dif-ference undoubtedly exists, and someportion of the effect can be so attributed.In fact, this factor probably accounts for thegreater magnitude of the MA superiorityin the present between-subjects design thanwas observed by Wickensand Kessel (1979)in a similar repeated-measures design. Inthat study, subjects in the AU detectioncondition had full access to an internalmodel of the system dynamics developedduring MA tracking on alternate trials.However, even in that investigation, inwhich tracking experience was the same


11/14

574 CHR ISTOPHER D . W I C K E N S AND COLIN KESSELwhen subjects were detecting in the AUor the MA mode, M A superiority was stillobtained.A third possib ility is that the requirementto track forced subjects to allocate moreresources to the tracking cursor itself in theMA mode, and this difference in the degreeof attention allocation induced the superiordetection. Results of two experimentalinvestigations, however, argue that the re-quirem ent to track is not by itself a sufficientcondition to produce greater atten tion to thespatial characteristics of the cu rsor. E phrathand Curry (1977), using a different sort offailure from that employed her e a spatialdeviation of the target off a linear trackobtained th e opposite results, showing AUsuperiority. This reversal of results is notinconsistent with the p resent findings. Thedeviation failure is one that does not entailany basic adaptation of MA control manip-ulation and therefore would not generateth e critical proprioceptive cues (W icken s &Kessel, 1979). Also, Klein and Posner(1974) observed that when presented aspatial-temporal pattern fo r later repro-duction, subjects were less accurate inthat reproduction when they tracked theoriginal pattern than when they passivelymonitored it .Although the above evidence countersarguments that M A superiority resultsexclusively from nonproprioceptive sources,further evidence can also b e marshaled tosupport th e assertion that the propriocep-tive channel is actively employed in bothtracking and in MA failure detection. T hus,Klein and Posner (1974) attributed thepoorer reproduction in their tracking con-dition in part to the forced division of at-tention between proprioceptive and visualchannels. A numbe r of investigations, fur-thermore, have found that tracking per-formance deteriorates as the quality ofproprioceptive feedback from th e controlstick is reduced (Bahrick, B enn ett, & Fitts,1955; Frost, 1972; Kessel & Wickens ,Note 2, Experiment 3). Curry and Ephrath(1976), furthermore, noted a direct loss infailure-detection performance as the qualityof proprioceptive feedback was similarlyattenuated.

Since the present results argue for theparallel processing of visual and propriocep-tive information in the manual mode, it isinstructive to consider these within theframework of the information-processingapproach to the visual-dominance phenom -enon, (e.g., Kelso, C ook, Olson, & Epstein,1975; Kelso & W allace, 1978; K lein , 1976;Posner, Nisson, & Klein, 1976). A visual-dominance interpretation would seeminglypredict that following a transition and subse-quent control adaptat ion, informationreaching a central decision mechanism fromthe visual sense would suppress that ar-riving along proprioceptive channels. Fur-thermore, the control adaptation to thechanged dynamics implemented by thesubject in the MA condition, if effective,should substantially reduce the magnitudeof visual error information concerning thetransition (i.e., the error signal shou ld be thesame as b efore transition) relative to visualerror produced by the nonadapting auto-pilot. The greater MA adaptation shouldthereby produce AU superiori ty. Thereason this effect was not obtained logicallyrelates to a further elaboration of the visual-dominance hypothesis suggested b y Klein(1976) and Kelso and Wallace (1978). Thiselaboration asserts that th e bias towardprocessing vision and suppressing proprio-ception will only b e evident when there isno reason to rely on the proprioceptivechannel to deliver task-relevant informa-tion. In the present paradigm the proprio-ceptive channel could produce relev ant anduseful information bearing on the occur-rence of transitions, and so the naturalvisual bias would not operate. In fact, K elsoand Wallace have argued that such con-ditions may induce a bias toward proprio-ception. Subjects in the present M A con-dition could therefore presumably benefitfrom some degree of bisensory facilitationto increase MA detection performance, afacilitation observed under similar non-biasing conditions b y Klein (1976) in a multi-mode choice reaction-time paradigm.The Structure of Resources in Detection

The present data suggest that the AUdetection processthe comparison of in-


12/14

PR OCESSING R E SO U R C E D E M A N D S 575coming visual data w ith an internal mo del ofth e normally functioning system and theapplication of appropriate decision rules toinitiate the responsedepends on re-sources residing in the reservoir associatedwith encoding, memory, and transforma-tions. The mental-arithmetic task alsodemand s these central-processing resources.In contrast, the critical task, whosedynamics and required transformations arethose of a relatively simple first-order sys-tem, does not extensively load these reser-voirs. I ts resource demands do not competeheavily with those of AU detection butapparently draw from the response-relatedreservoir. This view is consistent with theresults of a series of investigations in whichtracking has been paired with secondarytasks that elicit evoked brain potentials(Isreal, C hesney, W ickens, & Donchin,1980; Isreal, W ickens, C hesne y, & Do nchin,1980; Isreal et al., 1979). In these investi-gations, the evoked potentials, assumed tobe independent of response-related factors,since no overt responses are required fortheir elicitation, are found to b e insensitiveto a variety of manipulations of trackingdifficulty but to reflect the demands of per-ceptual cognitive task s.In MA detection, the processing of theresponse-related proprioceptive informa-tion channel involved in the early phasesof information integration is disrupted bythe demand for response resources of thecritical ta sk, b ut the processing is unaffectedby the central-processing demands ofmental arithmetic. Conversely, AU detec-tion, dependent o nly on visual information,is unaffected by response-related critical-task demands but is disrupted by thedemands on the central-processing pool im -posed b y mental arithmetic. Both detectionmodes involve a common central-process-ing/decision-making stage, which initiatesthe discrete manual response. H ow ev er,the resource demands of these processesare presumably reasonably light and do notprovide a source of task interference.In the present interpretation, it will benoted that the concept of processing stageshas been uncoupled from that of processingreservoirs. This uncoupling is portrayed in

Encoding Responding

PROCESSINGSTAGES

TASKRESOURCEDEMANDS

RESOURCERESERVOIRS

Figure 3. Representation of resource pools an d pro-cessing stages utilized b y dynamic system (MA ) an dautopilot (AU ) detection and by the two loading tasks.(Height of the task bars indicates demand levels.)Figure 3 and suggests that two separatestages of processing (perceptual encodingand central processing) b oth rely on a com-mon reservoir of resources. Some evidencefor the commonality of resource demandsof those stages is provided by investigationsin which detection and memory tasks arefound to exhibit considerable mutual inter-ference and to show corresponding per-formance-difficulty trade-offs (e.g., Shul-man & Greenberg, 1971; see Wickens , 1980,for a review). In contrast, the responsestage draws resources from its separatereservoir. The critical task relies on theresponse reservoir as does M A detection,whereas the m ental-arithm etic task and A Udetection depend on central-processing/encoding resources. These relations aredepicted in Figure 3.The interpretation presented above iscertainly not definitive, and alternate ex-planations of the data could be offered.However, it should be emphasized that theframework within which this interpretationis proposedthe concept of mu ltiple reser-voirs that supply stages with processingresourcesis consistent with many find-ings in the experimental literature that aresummarized by W icken s (1980), and in amore general way with theories proposed b yKinsbourne and Hicks (1978), Navon and


13/14


14/14

PROCESSING RESOURCE DEMANDS 577dominance: An information processing accountof its origins and significance. Psychological Re-view, 1976,83, 157-171.Rasmussen, J. Reflections on the concept of operatorworkload. In N. Moray (Ed.), Mental workload:Its theory and measurement . New York: PlenumPress, 1979.Roediger, H. L., Knight , J. L., & Kantowitz, B. H.Inferring decay in the shortterm m emo ry: The issueof capacity. Memory & Cognition, 1977, J, 167-176.Rouse, W. R., & Gopher, D. Estimation and controltheory: Application to modeling human behavior.Human Factors, 1977, 19 , 315-331.Sanders, A. Some remarks on mental load. In N.Moray (Ed.), Mental workload: Its theory andmeasurement. New York: Plenum Press, 1979.

Sheridan, T. B. On how often the supervisor shouldsample. IEEE Transactions in System Science an dCybernetics, 1970, SSC-6, 140-145.Shulman, H. G., & Greenberg, S. N. Perceptualdeficit due to division of attention between per-ception and memory. Journal of ExperimentalPsychology, 1971,88, 171-176.Smallwood, R. D. Internal models and the humaninstrument monitor. IEEE Transactions on HumanFactors in Electronics, 1967, HFE-8, 181-187.

Toates, F. M. Control theory in biology and experi-mental psychology. London: Hutchinson, 1975.Watson, D. S., & Nichols, T. L. Detectability ofauditory signals presented without defined observa-tion intervals. Journal of the Acoustic Society ofAmerica, 1976,59, 655-668.Wickens, C. D. The effects of divided attention oninformation processing in manual tracking. Journalof Experimental Psychology: Human Perceptionan d Performance, 1976,2, 1-17.Wickens, C. D. The structure of processing resources.In R. Nickerson (Ed.), Attention and performanceVIII. New York: Erlbaum, 1980.Wickens, C . D., & Kessel, C. The effect of participa-tory mode and task workload on the detection ofdynamic system failures. IEEE Transactions onSystem, Man and Cybernetics, 1979,13 , 24-31.Wicklegren, W. Speed accuracy tradeoff and infor-mation processing dynamics. Acta Psychologica,1977, 41 , 67-85.Young, L. On adaptive manual control. IEEE Trans-actions on Man Machine Systems 1969, MMS-10,292-331.

Received June 18, 1979

Wickens and Kessel 1980

Documents