IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-9, NO. 1, JANUARY 1979

The Effects of Participatory Mode and

Task Workload on the Detectionof Dynamic System Failures

CHRISTOPHER D. WICKENS AND COLIN KESSEL

Abstract-The ability of operators to detect step changes in theorder of control dynamics is investigated as a joint function of a)participatory mode, whether subjects are actively controlling thosedynamics or are monitoring an autopilot controlling them, and b)concurrent task workload. Five subjects either tracked or monitoredthe system dynamics on a two-dimensional pursuit display undersingle task conditions and concurrently with a "subcritical" trackingtask at two difficulty levels Detection performance was faster andonly slightly less accurate in the manual as opposed to the autopilotmode. Performance in each mode was derogated by the concurrenttracking requirement, but not by increases in loading task difficulty.Further analysis indicated that manual superiority was attributableto the additional proprioceptive information resulting from operator-control adaptation to the system change.

INTRODUCTION

OVER THE past decade, the aviation industry haswitnessed a gradual change in the role of the pilot in

the cockpit. As flight dynamics have become increasinglycomplex and as computer technology has advanced ac-cordingly, many traditional pilot functions have beenreplaced by on-board computers, and in some instances thepilot is no more than a supervisor [1] or monitor ofautomatically controlled functions. One task, however, thatremains of critical importance to the operator of anyaviation system, whether he is removed from the controlloop or not, is that of monitoring all facets of aircraftperformance for the occurrence of failures or malfunctions,for example, the loss of stability augnentation, or loss ofpower to an engine. The relatively low frequency of occur-rence of such events does not diminish the importanceof failure monitoring and detection because the con-

sequences of an undetected malfunction or one that isdetected after an unnecessary delay can be disastrous,potentially resulting in the loss of the aircraft or of humanlife.Young [2] has argued strongly on the basis of his findings

that the operator is more sensitive to system malfunctions as

an active participant in the control loop than as a passivemonitor. In his experiment, subjects were required to detectvarious step changes in system order and gain. Conditionswere compared in which the subject was an active controllerand a passive monitor (who was observing the compensa-

Manuscript received October 31, 1977; revised April 20, 1978, and Sep-tember 11, 1978. This work was supported by the Life Science Program,Air Force Office of Scientific Research, under Contract F44620-76-C-0009.The authors are with the Department of Psychology, University of

Illinois, Urbana-Champaign, IL 61820.

tory display produced by another active controller). Underthese circumstances, detection latencies were two to fivetimes greater for the monitor than the controller. Contraryconclusions, however, were drawn from investigations byVreuls [3] and by Ephrath and Curry [4]. The latter studyinvestigated failure detection performance in a two-dimensional simulated landing task as a function of partici-patory mode. The "failures," which in this case weredeviations introduced into the flight path rather thanchanges in system dynamics, could occur in either the lateralor longitudinal control axis. Under different conditionssubjects were either in control when a failure occurred orwere monitoring a nonadaptive autopilot in control of thatchannel. The nonfailed channel could also be either con-trolled or monitored. Ephrath and Curry's results indicateda clear superiority for detection on the monitored asopposed to the controlled dimension, both in terms of thesmaller number of missed failures and of the shorter detec-tion latency. This difference was attributed jointly to theincreased level of workload and to the allocation of atten-tion away from critical sources of failure-related informa-tion in controlling condition.

Obviously, in many respects the studies of Young and ofEphrath and Curry are not comparable. Young employedsingle-axis tracking with changes in system dynamics, whileEphrath and Curry employed dual-axis simulator controlwith "deviation" failures. In addition, the monitoring condi-tions were different in the two experiments, since the "auto-pilot" monitored by subjects in Young's study, beinganother human controller, was thereby capable of adaptingto the change in dynamics. The autopilot employed byEphrath and Curry, however, was not adaptive. The study ofEphrath and Curry also employed a more realistic flight taskin the simulator, with a considerably greater level of taskworkload. In light of the many differences between thestudies, it is not surprising that their conclusions differed aswell. The present investigation represents an attempt to shedfurther light both on the differences between failure detec-tion in the participatory modes and the different conclusionsof the two experiments.

Descriptive Analysis of the Failure Detection Process

The detection of a failure or change in the characteristicsof a dynamic system requires that the operator have avail-able two basic elements: 1) an internal representation of thestate of the normally operating system-the expected value

0018-9472/79/0100-0024$00.75 ( 1979 IEEE

24

25WICKENS AND KESSEL: DYNAMIC SYSTEM FAILURES

of state variables and their expected variability [5]-[9], and2) a channel, or set of channels, of information used toestimate the current state of the system. Noisy informationconcerning the system state is integrated over time andfailures are detected by a decision mechanism when thecurrent representation is assessed to be sufficiently deviantfrom the representation of normal operation to warrant a

response. The decision process may be assumed to involvethe application of some statistical decision rule [7]. Giventhat a failure has occurred, the diagnostic value of the noisyinformation will increase with the number of informationsources and with integration time, as long as memory for thestandard of normal operation remains salient. Providedmemory is salient then, failure decisions should becomemore prevalent at longer latencies. However, if sufficienttime has lapsed with the post-failure dynamics, the internalmodel of normal operation itself should begin to reflectpartially the new dynamics, and the "strength" of thedifference signal that is integrated thereby becomes at-tenuated. Thus while integrated information grows over

time, the diagnostic value of that information will eventuallydecline, dictating that detection accuracy (or number ofdetections) will not be monotonically increasing withlatency but rather will reflect this trade-off, achieving a

maximum at an intermediate latency. Within the frameworkof this representation, it is possible to identify areas ofcontrast between processing in the two modes that couldaffect detection performance. These areas are detailed as

follows.Consistency ofthe Internal Model ofDynamics: Detection

performance will clearly be influenced by the consistency ofthe operator's representation of the controlled system. Thisconsistency corresponds to the variability of expectedsystem outputs in response to any given input. If thedynamics of the system are highly familiar to the operatorand model consistency is great, then deviations from thesedynamics should be rapidly detected. It is argued here thatan active controller, having the opportunity to clearlydifferentiate his own input to the system from disturbancesacting upon it, will be better able to construct a consistentinternal model and so be facilitated in detecting dynamicschanges.

Information Channels: While monitoring, the operatorhas available information concerning the occurrence of a

failure only from the visual display. These visual quantities,either directly perceived or derived, correspond to thesystem state and error and their respective time derivatives.The controller on the other hand has available, in additionto visual display information, a second parallel channel ofpotentially relevant information in the form of the pro-

prioceptive input generated by control manipulation eman-

ating from both joint position and muscle-tension relatedreceptors [10]. Although control manipulation cannotdirectly reflect the occurrence of failures (except as failuresinitiate mechanical feedback from the control itself), it willdo so indirectly to the extent that any compensatory adapta-tion that the operator initiates to a system change will bereflected in a change in his response characteristics (meancontrol position, velocity, or acceleration) and a corre-

sponding change in the characteristics of the operator'sopen-loop transfer function [2]. When controlling then,these proprioceptive channels will be available to the detec-tion system to supplement the visual channels that areavailable in both monitoring and controlling. Assuming thatthe two sensory channels available to the controller convey-ing failure information are perturbed by uncorrelated noise,the diagnostic value of the joint information available to thedecision mechanism should exceed that of the informationavailable to the monitor [11]. Curry and Ephrath [12]providing support for the role ofproprioception in detectionhave found that when the quality ofproprioceptive feedbackis reduced by employing an isotonic control, failure detec-tion performance deteriorates accordingly.An important qualification of the above statement is that

information concerning a failure occurrence in the pro-prioceptive and in the visual channels are not independent ofeach other but are interrelated through the process ofcontrol adaptation to the post-failure dynamics. If adapta-tion to the failure is rapid and complete, as may occur forexample in response to shifts in system gain [2], the obtaineddistribution of error following the change would show littleor no alteration from that characterizing the normal operat-ing state, while a change would be manifest in the character-istics of the control response. Conversely, failure to initiateany adaptive control would leave unchanged the pro-prioceptive input, while altering the nature of the displayederror distribution. The importance of this point is describedbelow.

Differential Sensitivity to Visual versus ProprioceptiveInformation: While the controller will be provided with atleast as many channels ofinformation as the monitor (one inthe absence of adaptation, two in its presence), it should beemphasized that the information along the two channels willbe a function not only of the extent ofadaptation, but also ofthe relative efficiency of detecting failure-related informa-tion along the visual versus proprioceptive channels. Thissecond factor is of potentially critical importance since thereis considerable experimental evidence that sensitivity toproprioceptive information is reduced relative to visualinformation, particularly when the two sources are availableat the same time and are conveying conflicting information,e.g., [13]-[15]. Such a conflict, in fact, describes precisely thesituation in which an operator has successfully adapted to achange in control dynamics. Under these circumstances, thevisual error channel is providing information describingnormal operation (since the appropriate gain, or lead-lagadjustment, has presumably been initiated to restore theoriginal open-loop transfer characteristics), while the lesssensitive kinesthetic channel conveys the information that achange has in fact been implemented. The predicted con-sequence of this conflict situation is that the operator will beless likely to detect the change than he would had noadaptation been achieved, the latter condition of courseproducing a visual signal equivalent to that of a nonadaptivemonitored autopilot. McDonnell [16], in fact, has notedanecdotally such instances in which successful adaptationhas been coupled with the failure to detect dynamic systemchanges.

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-9, NO. 1] JANUARY 1979

Workload Differences: A second characteristic of themanual control mode that predicts a reduced sensitivity tothe occurrence of failures relates to the greater workloadimposed by tracking than by monitoring. Numerousexamples may be cited from behavioral literature thatdemonstrate the attention demands of purely perceptualtasks such as monitoring to be less than those of tasks suchas tracking in which a requirement for the selection andexecution of responses is also imposed [17], [18]. Thisfinding is verified as well in a direct comparison of control-ling versus autopilot monitoring in the simulator [4], [19]. Inthe framework of the present analysis, if monitoring for andresponding to failures is regarded as a "task" separate fromtracking, then since the operator's attentional resources arelimited, the greater workload demands imposed in thecontrol mode than in the monitoring mode would predictpoorer performance on the added "task" of failure detectionfor the operator in the control loop.

Hypothesis Testing: One final difference between modesthat cannot be overlooked relates to the operator's ability inthe controlling mode to "test" the dynamics when thesuspicion of a failure arises, by injecting artificial signals intothe system and observing the response. This hypothesistesting is an option which is normally unavailable to theautopilot monitor.

SummaryThe implications of the preceding analysis are complex. In

summarizing, three attributes of the controlling mode maybe identified that would seemingly facilitate failure detec-tion: a smaller variability ofthe internal model of the system,the option of hypothesis testing, and a greater number ofchannels available upon which to base failure-detectiondecisions. At the same time, the latter advantage may bemitigated to the extent that a) adaptation takes placereducing the strength of a visual error signal and b) pro-prioceptive sensitivity is less than visual. In comparison, themonitoring mode is also characterized by two attributes thatcould facilitate detections: a greater "strength" of the visualsignal (ifadaptation by an autopilot does not take place) anda lower level of workload. Clearly this interplay of factors is

sufficiently complex to preclude precise predictions concern-ing the superiority of one mode over the other. The aboverepresentation, however, facilitates a clearer identification ofthe nature of the failure-detection task and allows predic-tions to be formulated concerning the differential effect ofvariables such as workload or control adaptation on detec-tion performance. The present study was conducted with theintent of clarifying the nature of the superiority relation in

failure detection between the two participatory modes.Failures were step increases in the second-order componentof system dynamics, thus simulating the loss of stabilityaugmentation in aircraft control. A question of specificinterest was whether the difference between the results ofEphrath and Curry and of Young could be attributed todifferences in concurrent task workload between the para-digms. For this reason task workload, the demand forprocessing resources imposed by a concurrent loading task,was manipulated orthogonally to participatory mode.

METHOD

SubjectsThe subjects were five right-handed male university stu-

dents enrolled in basic flight training courses at the Instituteof Aviation. Subjects were paid at a rate of $2.50 per hour.

Apparatus

The basic experimental equipment included a 7.5 x 10 cmHewlett Packard Model 1300 CRT display, a spring-centered dual-axis tracking hand control with an index-finger trigger operated witlh the right hand, and aspring-loaded finger controller operated with the left. ARaytheon 704 16-bit digital computer with 24 k memory andA!D, D/A interfacing was used both to generate inputs to thetracking display and to process responses of the subjects.The subject was seated on a chair with two arm rests, one forthe tracking hand controller and one for the side-task fingercontroller. The subject's eyes were approximately 112 cmfrom the CRT display so that the display subtended a visualangle of 3.4°.

Trackin1g Tasks: The primary pursuit-tracking taskrequired the subject to match the position of a cursor withthat of a target which followed a semipredictable two-dimensional path across the display. The target's path wasdetermined by the summation of two nonharmonicallyrelated sinusoids (0.05 and 0.08 Hz) along each axis with aphase offset between the axes. The position of the followingcursor was controlled jointly by the subject's control re-sponse and by a band-limited forcing function with a cutofffrequency of 0.32 Hz for both axes. Thus the two inputs tothe system were well differentiated in terms of predictability,bandwidth, and locus of effect (target versus cursor). Thecontrol dynamics of the tracking task were of the formYI= (1 - a),`s + (a/s2) for eaclh axis, where o was the var-iable parameter used to introduce chaniges in the systemdynamics. These changes, or simulated failures, were in-troduced by step changes in the acceleration constant a froma normal value of 0.3, a mixed velocity and accelerationsystem with a high weighting on the velocity component, tox = 0.9, a system that approximates pure second-orderdynamics and requires the operator to generate consider-able lead in order to maintain stable performance.As the loading task, the critical task [20], was employed.

This was displayed horizontally at the bottom of the screenand required the subject to apply force to the finger controlin a left-right direction to keep the unstable error cursorcentered on the display. The value of the instability constantA in the dynamics Y, = ki (s -) was set at a constantsubcritical value. Two values (/,. 0.5 and i. - 1.0) were

employed on different dual task trials.

Experimenital Task

Subjects participated in five experimental sessions ofwhich the first two were devoted entirely to practice on thetracking and detection tasks and the last thlree used to

generate the experimental data. During the first practiceday, the subject performed only the two-dimensional pursuittracking task. In the manual (MA) coniditioni the subject

26

WICKENS AND KESSEL: DYNAMIC SYSTEM FAILURES

performed the tracking manually, while in the autopilot(AU) condition, his role in the control loop was replaced bysimulated autopilot control dynamics consisting of a puregain and effective time delay. The open-loop gain was set at aconstant value for all subjects, and the autopilot time delaywas increased from zero for each subject to a value at whichthe rms error was equivalent to that subject's performance inthe MA condition. This value of time delay was maintainedthroughout the rest of the experiment. Each trial, MA orAU, lasted 150 s.To provide some experience with the failed condition (i.e.,

the higher acceleration in the control dynamics), the subjectreceived two trials (one AU and one MA) in which hetracked (or viewed the autopilot tracking) only the faileddynamics. Two demonstration trials were then presented inwhich the subject tracked in the regular condition, but theonset of each failure, was cued by the presentation ofan F onthe screen. The subject was instructed to press the trigger toreturn the system to normal only upon the detection of thenature of the change. This training period was then followedby eight regular detection trials (4 AU, 4 MA in alternatingorder). Each trial contained either four or six failures so thata total of 20 failures were presented in each mode.The presentation of the failure was generated by an

algorithm that assured random intervals between presenta-tions and allowed the subject sufficient time to establishbaseline tracking performance before the onset of the nextchange. Task logic also ensured that changes would only beintroduced when system error was below a criterion value.In the absence of this latter precaution, changes wouldsometimes introduce obvious "jumps" in cursor position.

During these detection trials, the detection decision wasrecorded by pressing the trigger on the control stick. Thisresponse presented a T on the screen and returned thesystem to normal operating conditions of the prefailuredynamics. If the subject failed to detect the change, thesystem returned to normal after 6 s via a 4 s ramp. The sixseconds were an interval within which it was assumed, on thebasis of pretest data, that responses would correspond todetected failures and not to false alarms. The subjects weretold to detect as many changes as possible as quickly aspossible.On the second day (dual-task training), the subject per-

formed the primary tracking task together with a side task,the critical task. After a refresher trial in the MA mode, thesubject received a series of training trials to practice the sidetask, first in the AU and then in the MA mode. Whenacceptable criteria were achieved in the critical task and MAtracking individually, the subject then carried out thesetasks together with the failure demonstrations, as describedabove.

Eight more experimental trials were then presented inwhich the subject performed all three tasks (tracking ormonitoring, critical task, and failure detection). Two trialswere presented in each mode at each level of critical taskdifficulty (A' = 0.5, 1.0). The subject was instructed to "do theside-stick task as efficiently and accurately as possible." Theinstructions, therefore, clearly defined the side task asthe loading task while allowing performance on the tracking

TABLE IWITHIN SUBJECT EXPERIMENTAL DESIGN (DAYS 3, 4, 5)

Participatory Mode

Auto (AU) Manual (MA)

Tracking 30 failares 30 failuresSingle

Task

tsackin: ana

asy Critical 30 failures 30 failures

TaskDua I

Ta ;k

racA ing and

iff icult 30 failures 30 failures

ritical Task

and detection tasks to fluctuate in response to covertchanges in available attentional resources. These instruc-tions were emphasized by providing subjects with trial bytrial feedback on critical task performance. In this manner,workload demands were experimentally manipulated, ratherthan being passively assessed.

Following the two training days, the final three days, usedto generate the data for experimental analysis followed theformat of Table I. The order of presentation of the 12experimental trials was counterbalanced across subjects andacross days within a subject. The task logic, instructions, andexperimental procedure was otherwise identical to that ondays 1 and 2.

ANALYSISDetection performance was assessed in terms of the

accuracy and latency of responses. Because accuracy is ajoint function of the proportion of failures detected and thenumber of false alarms (detection responses in the absence ofa failure), signal detection theory analysis was employed[21]. However, a modification of the conventional signaldetection analysis procedure was adopted because in theparadigm employed subjects were not provided with aclearly defined response interval. According to thisprocedure, known as the method of free response [22], anassumption is made that any response occurring within agiven interval following a failure is a hit a responsetriggered by the failure signal. Any other response is a falsealarm. Data collected in pretests were used to define the hitinterval as 6 s, and the measure P (HIT) was simply thenumber of detection responses falling within the intervaldivided by the total number of intervals. Following theprocedures outlined by Watson and Nichols [22], the re-maining duration of the trial was subdivided into false alarmintervals of duration equivalent to the 6-s hit intervals.The measure P(FA) was computed as the number of falsealarms divided by the total number of false-alarm intervals.For a sensitivity measure, the area under the ROC

[A (ROC)] was employed rather than d' because the formermeasure is more robust to violation of distribution assump-tions and to the small number ofsignals employed here [21].

27

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-9, NO. 1, JANl!ARY 1979

TABLE IIDETECTION DATA

15A AU

SEubjec t Sing le task A .5 1 1.0 Sin01c ta - .5 - - -

SI P(11) .40 .267 .267 .7 , .700 .4'7

1P(FA) .071 .060 .060 .130) .100 .239

s2 P(H) .7h) 7 .53, .567 .83i .533 .533

P(FA) .050 .150 .120 .120 .190 .150

S3 P(i1) .4o 7 .4-33 .567 .5 3 .458 .423

iG ) :, 7 (0 .2(1)220 . 2 '( .164 , i S.

S4 P(O1) .267 .375 .200 .533 .700 .429

P(FA) .110 .1 3 g .1/( .120 .110 .1/

s5 P(NH) .500 .1 i .6/ .613 .816 .

P(FA) .0/ .0X', .0/3 .058 .U'() .0/0

A(ROC) is equivalent to the area in the unit ROC spacebelow and to the right of the ROC curve connecting thesingle P(H), P(FA) observation. This measure produces ascore varying from 0.5 (chance accuracy) to 1.0 (perfectperformance).' Values of this measure were taken fromtables in McNicol [23].

Tracking measures of vector error, vector control posi-tion, and critical task error were sampled every 60 ms andstored on digital tape for later data analysis. Error andcontrol position were also differentiated to obtain theirrespective velocity values. In addition, on a fourth channel,the occurrences of failures and responses were recorded. Atthe end ofeach trial, the rms vector error on the primary taskand rms error on the critical task (if performed) werecomputed.

RESULTS

Detection PerformanceTable II presents the average hit and false-alarm probabi-

lities for each subject in each of the six conditions, andfollowing the procedure outlined above, these were con-verted to A(ROC) measures. Performance in the differentconditions plotted as a joint observation of detection accur-acy [A(ROC)], and latency is shown in Figs. 1 and 2. In Fig.1, the heavy vectors portray the effect of adding the loadingtask on the mean performance of all subjects in the speed-accuracy space. The beginning of the vector thus corre-

' The A(ROC) representation does not allow a separate assessment ofresponse bias. This dimension of performance was not examined in thepresent study because of the greater interest in detection sensitivity andthe dependence of any response bias measure upon assumptions concern-ing the penalty and values associated withl different response outcomes.Since no explicit payoff structure was imposed upon the subject's re-

sponses. these biases may have varied considerably.

sponds to the bivariate mean of the five individual subjectdata points assessed under single task conditions. The vectorend (arrow) corresponds to the bivariate mean of theaverage performance measures across the two critical taskconditions. The lighter vectors portray the data of the fiveindividual subjects. Because of the importance of viewingindividual subject data, the brackets below represent theaverage magnitude, across subjects of + I standard errorconfidence estimates along both the latency and accuracyaxes.2 By this representation, it is possible to view simultan-eously the trend of a subject's behavior in the speed-accuracy space as the loading task is required, the extent towhich the trend typifies the behavior of all subjects, andthrough the confidence brackets, the reliability of the trendsshown by individual subjects. Since considerable exper-imental literature indicates that adding a concurrent taskand increasing its difficulty may have qualitatively differenteffects on primary task performance [24], [25], a separatedepiction of the effects of increasing critical task difficulty isshown in Fig. 2. Here the vector portrays the effect ofincreasing the instability constant A from 0.5 (vector begin-ning) to 1.0 (vector end).The rationale for the joint speed-accuracy representation

of Figs. 1 and 2 is that these two variables represent differentmanifestations of an underlying performance metric. In anyeffort to compare "performance" across conditions, the jointimplications of speed and accuracy must be taken into

2 For the accuracy data, standard error of proportion scores werecomputed for hit and false-alarm probabilities by the formulaSp= \/[P(1 - P)]/N Standard-error confidence brackets were thencomputed for the accuracy measure A(ROC) by determining the maxi-mum and minimum accuracy values obtainable when P(Hit) and P(Falsealarm) were within one standard error of their respective estimated values,e.g., maximum accuracy: P(H) + lSP, P(FA) - ISP; minimum accuracy.P(H) lSp P(FA) + 1SP

28

WICKENS AND KESSEL: DYNAMIC SYSTEM FAILURES

-

Fig. 1. Eacy and(thin ve

Fig. 2. E

account

high acc

latency

context:lower e

model, 1until a daccurac'

criteriorjoint reI

In Figrepreser

Single Dual mance is in the lower right regions. In an orthogonalTask Task i ih na

(Collapsed) direction, shifts in bias for speed versus accuracy correspondMA - to movement between the lower left (speed bias) and theAU * upper right (accuracy bias) in the space. These shifts may be2 g inferred to relate to variations in setting ofthe decisionmak-

Tt2 ing criterion.Undoubtedly, the most noteworthy effect in Fig. 1 is the

increase in response latency from the MA to the AU_ _ _ _ >4 > . conditions (t = 5.84, p < 0.001). While this increase in

latency for the AU mode is consistent for all subjects, it isaccompanied by an increase in response accuracy that is less

- / 3 ~ ~ ~pronounced and is only evident in both single and dual-task4X 3 conditions for three ofthe five subjects (SI, S4, and S5). Thus

the prevailing trend induced by shifting from the MA to AUmode appears to be a shift in bias to the upper right in thespeed accuracy space, towards slightly more accurate, butconsiderably slower detection.

40

I0I The change in performance induced by the requirement to

2.0 3.0 4 perform the additional critical task (Fig. 1) is in the directionLatency (sec) of poorer performance increased latency and decreased

ffects of participatory mode and critical task on detection accur- . . 'latency. Mean subject trend (heavy vectors). Individual subjects accuracy. This is a trend predictable from the assumption ofctors) are identified by subject numbers adjacent to data points. limited operator processing resources which, when diverted

to the loading task, derogate detection performance. Themean vectors indicate that this trend is of about the same

Easy Difficult magnitude and in the same direction for both the AU andX = s x= 1o MA modes, although it is considerably less consistent acrossMA o subjects in the MA condition, with two subjects (S3 and S5)AU 4 - - showing vectors that do not run in the predicted direction.

--- I- *4 S3 shows an accuracy increase, and S5 shows a latencydecrease with critical task performance.

In Fig. 2, presenting the effect of increasing critical taskdifficulty on detection performance, the mean trend in both

\ x * 3 modes shows a decrease in accuracy. However, this is only4 r consistent across subjects in the AU mode (with minor

exceptions of S2 who shows a minimal accuracy increase)./ ~s,l Furthermore, in both modes the increase in critical task

difficulty appears to lead to a slight decrease in averageresponse latency. Subjects S5 (MA mode only) and SI areexceptions here.While the trends in neither data set of Fig. 2 are striking,

A 1 I the finding of interest is perhaps the very fact that trends are3.0 4.0 not pronounced when they might otherwise have been

Latency (sec) expected with increased loading task difficulty. The reasonffects of participatory mode and critical task difficulty on detec- for this expectation is evident from Fig. 3, which presents the

tion accuracy and latency. effects of the workload manipulation on tracking perfor-mance. It can be seen here in the MA mode that tracking

[26], [27]. For example, a condition that produces a performance deteriorates equally as the critical task is added,uracy ofresponding might do so at such a prolonged (t = 3.05, p < 0.01) and as its difficulty increases (t = 1.71,that the utility of that decision in a real-world p < 0.05) (the manipulations corresponding to thoseis less than that of a more rapid decision with slightly portrayed in Figs. 1 and 2, respectively). Even though therexpected accuracy. Furthermore, the underlying is a slight increase in critical task error with i, it is safe toproposing that information is integrated over time conclude that increasing critical task difficulty leaves fewerlecision criterion is reached, suggests that speed and of the operator's limited processing resources available toy may be "traded off' by manipulating the decision allocate to the tracking task. Yet in the MA mode, thisn. This trade-off presents another justification for the diminuition of resources does not appear to reduce detec-presentation. tion performance in any marked fashion for other than S5.gs. 1 and 2, "good" performance (fast and accurate) is Similarly in the AU mode, the trend of detection perfor-ited in the upper left regions, while poor perfor- mance with increasing A cannot be described as an unambi-

-

u

0x

cl

U

0

I..

M

8

.7

.6

U0

0C-

c)eu

9

.8

.7

.6

.5

29

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-9, NO. 1. JANt ARY 1979

30F

0

LUCO) .201-

10

S2

Im

U--_- - 40

.,-*CriNicol-- ~~Task

L

Single A = 5 A = /0Task -Cond,tion_

(a)

SI

L .. 4-.

S4

S3

l- ,

(b)Fig. 3. Effects of critical task difficulty on performance of primary track-

ing and of Critical Task. (a) Subject means. (b) Individual subjects.

guous decrease in performance, but rather as a shift in bias tofaster, but less accurate, responding (Fig. 2).

Information Utilized in Detection

In the introduction, three hypotheses were proposed topredict why detection might be superior in the MA over theAU mode, and two were proposed predicting AU superior-

ity. The results presented in Figs. 1 and 2 suggest thatmanual detection was generally superior, and the data were

analyzed in detail to determine their consistency with thesecond hypothesis proposed for manual superiority i.e.,

the role of the added proprioceptive channel. In this endea-vor, three analysis techniques were pursued to identifythe cues employed and to provide insight into the nature ofthe detection process.

1) Ensemble averages of display and control variableswere constructed primarily to determine the existence offailure "signals" that might have been employed in detec-tion. These signals are time varying characteristics of thevariables, time-locked to failures, and increasing during thepost-failure interval (Fig. 4). Separate averages were con-

structed for hit and miss trials in each condition.

2) Multiple regression techniques were employed todetermine what characteristics of the signial and responsewere the best predictors of detection latency. As predictorvariables error, error velocity, control velocity, anld cursorvelocity were sampled at the instant of failure, at 0.6. 1-2. 1.8.2.4, 3.6, and 4.8 s after the failure. The resuilts of theregression analysis are presented in Table llI. The three bestpredictor variables (those jointly accounting for tlhe greatestproportion of latency variance) are presented f'or the MAand AU single-task conditions. Associated with each is tihepartial correlation coefficient between the designated var-iable and latency.

3) The distribution of detection responise latencies wasexamined to assess the accumulation of failure-relatedevidence by the decision center. The results of' these tlhreetechniques will be discussed as they bear upon the questionof the cues utilized for detection.The assumption that AU detection is based upon the

error information is borne out by the single task ensembleaverages presented in Fig. 4. Clearly a transient increase indisplayed error is produced by the failure (information isavailable to the decision center), and the difference betweendetected failures (solid lines) and undetected ones (dottedlines) is consistent with the view that, of the randomlydistributed set of error profiles following failures, those thatmanifest a smaller increase tended to be missed. Thesetherefore generated the lower ensemble. This anialysis iscorroborated by the multiple regression data (Table Ill).The negative value for the best predictor of AU latencysuggests that larger error signals at 0.6 s latency are asso-ciated with faster responses. Similarly the second predictorvariable, error velocity, is also associated with latency insuch a way as to suggest that increases in its magnitude serveas a signal to shorten detection latency.Turning to the MA condition in Fig. 4, the role played by

displayed error is again evident. Error increases followingthe failure and therefore is available as a signal (altlhouglh itapparently does not increase differentially between hit-and-miss trials.) Furthermore from Table III the best predictor ofresponse latency is again the error variable, this timesampled at 1.2 s post-failure.The fact that the increase in the average hit and miss error

traces is greater in the AU than in MA condition suggeststhat in the MA mode the operator is performing some sort ofcontrol adaptation to the new post-failure plant dynamicsan adaptation directed to bring error to its prefailure level.According to predictions of the crossover model [28], theincrease in system order produced by the failure requires theoperator to develop greater lead, differentiate the errorvalue, and produce a response velocity of higher averagevalue. Thus to the extent that adaptatioin is carried out,control velocity should increase. Furthermore, even if adap-tation is not the linear response predicted by the crossovermodel, but represents instead a time-optimal bang-bangresponse [29], the later strategy should still produce anincreased control velocity.

In the data presented in Fig. 4, an increase in controlvelocity is visible following the failure, supporting the view

Conditione* MA

e- UAU

Primary Task

-1 mr- - I= I

30

.q s

eo-.it- - -.m

WICKENS AND KESSEL: DYNAMIC SYSTEM FAILURES 31

Detected Failures- - Missed Failures

AU

-~~~~ 0~~t

0-00~ ~ w IIJ~

/ ~~~~~~~~~~4..p 05

U).0

_ _ _ _ _

00 06 1.2 18 24 30 36 42

Latency After Failure (sec)

0

o

0

.0

< A , i

VAZI

MA

/

00 06 1.2 18 24 3.0 36 42

Latency After Failure (sec)

(a)

I I IIl

0 0 0.6 1.2 1.8 2.4 3.0 3.6 4.2

Latency After Failure (sec)(b)

Fig. 4. (a) Ensemble averages of error. (b) Control velocity. For single task AU (error only) and MA conditions. Verticalarrow indicates mean detection latency. Ordinate scales are in arbitrary units, but are same for AU and MA error averages.

TABLE IIIMULTIPLE REGRESSION ON RESPONSE LATENCY CONDITION

MA AU

Variable Partial Variable PartialName Correlation Name Correlation

Order ofPredictionVariable

I Error -.310 Error -.3571.2 sec. .6 sec.

2 Corntrol -.170 Error -.309Vel. 0.6 Velocity

.6 sec.

3 Error - .251 Error - .213Velocity Velocity1.2 sec. 2.4 sec.

Predictor variables were excluded if they occurred at latencies equal to or greater than the mean detection latencies.

0

w

0.5

.0 /

/

r/

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICS, VOL. SMC-9, NO. 1, JANIJARY 1979

that adaptation was carried out. This observation supportsthe hypothesis that MA superiority is based upon the addedproprioceptive channel of information, since the figureindicates clearly that this information (manifest in theincreasing control velocity) is available prior to detection asa channel to the decision center.3 A further corroboration ofthe use of this channel is found in the multiple regressionanalysis that identifies control velocity as the second predic-tor of response latency. The level of correlation, althoughmodest, is found to be consistently negative with latency, atthe 0.6- and 1.2-s post-failure time points.

Latenicy Distribhttion: The ensemble average and multipleregression analysis of the single-task data suggest that morerapid MA detection may be attributed to the adaptation-related proprioceptive information channel that becomesavailable to the decision center within 1-2 s following thefailure occurrence. This interpretation receives further sup-port from an analysis of the distribution of response laten-cies. In the current data, these distributions for all MAconditions were highly skewed in a positive direction, whilethose of the AU conditions were approximately symmetri-cal. The latency distributions were transformed to cumula-tive probability distributions portraying the relative numberor probability of failures detected, as a function of latencyafter failure (Fig. 5). Lappin [30] has argued that a similarrepresentation of his reaction time data the latency operat-ing characteristic or cumulative accuracy function --mayprovide evidence bearing upon the time-dependentprocesses involved in detection: the integration of evidenceover time.

Following Lappin's approach, it is argued here that thedata of Fig. 5 may be interpreted as follows. Since detectionlatencies are assumed to reflect, in part the instant at whichthe sampled evidence excedes the decision criterion, theextent to which accuracy increases as the criterion moves tolonger latencies (through between and within subject varia-tion) is a reflection of the rate of accumulation by thedecision center of failure-related information in the post-failure interval. Thus the slope of a function of relativeaccuracy versus latency represents the rate at which percep-tual evidence becomes available, while the level of thefunction or intercept represents the overall quality of thatinformation. According to this interpretation, three impor-tant characteristics are evident concerning the single anddual (e - 1.0) task data of Fig. 5.

1) Both of the MA functions indicate the presence of a

distinct discontinuity in the rate of accumulation ofevidence, this discontinuity occurring at approximately1- 1.5 s post-failure. In a manner consistent with the earlierdiscussion, it may be argued that the steeper early growthrate reflects the added availability of the proprioceptivesignal during the initial 1-2 s of adaptation, while the

That there is a difference between the hit-and-miss response velocityprofile of Fig. 4 indicates that response-related (proprioceptive) informa-tion is employed in detection. However, it is not apparent why the hit trace

of Fig. 4 shows initially a lower response velocity. It is possible that thisdifference reflects the initial sluggish response of the system and thereforethe initial low-velocity control response required to nullify the error.

When the low velocity response is pronounced. detection is facilitated.

0uC)

5:)

Qi

0

0~Q)

Z3U)

.01

MA-

S *

AU --UarU -_ -r

..

l

Latency after Failure(seconds)

Fig. 5. Cumulative probability distribution of detection latencies forsingle and dual (A = 1.0) task AtJ and MA conditions.

shallower slope following represents the integration ofevidence from the visually displayed error signal.

2) The two AU traces, while not strictly linear, fail to showthe abrupt discontinuity of the MA conditions, and thusseemingly represent a uniform underlying process. Thisprocess, presumably the integration of displayed visualinformation accumulates evidence at a faster rate (steeperslope) than is evident in the later visual portion of the MAmode. This interpretation is consistent with the smallerdisplay error profiles in the MA as compared to the AUconditions (Fig. 4) and with the assumption that anyadaptation in the MA condition should in fact reduce themagnitude of the visual error signal. With less informationtherefore available, the integration of this information willproceed at a slower rate.

3) In both MA and AU modes, the data of the dual-taskcondition lie below but closely parallel to the correspondingsingle-task value (intercept shift). This suggests that criticaltask performance, while affecting the overall quality ofperceptual data (or the delay in response initiation) in

identical manner for both modes, does not affect the rate ofits sampling or acquisition.

Role of Workload: The concept of workload ---the task-imposed demand for the limited processing resources of theoperator is relevant both as a potential source of AUdetection superiority, proposed in the introduction (but notshown by the data), and as it concerns the effect on detectionof the added requirement of critical task performance.Concerning the first issue, it is appropriate to ask why theadded workload of controlling the primary tracking taskapparently did not hinder the detection of failures in thistask, relative to the AU mode when this controlling functionwas not required and the subject's on7ly task was to monitorthe visual display for failures. Two answers may be provided.I) The role of both error signal and proprioceptive channelsdemonstrated in MA detection suggests that the very samemental operations that might on the one lhand be argued toincrease the competition for resources with the detection

32

33WICKENS AND KESSEL: DYNAMIC SYSTEM FAILURES

requirement are also the same ones that are integrallyinvolved in theMA detection process. These operations thenmay function in cooperation with the detection process,rather than in competition for the resources upon whichdetection depends. 2) A different accounting for the absenceof a workload deterioration effect can be proposed in termsof the nature of the processing resources themselves. Whilethe resources involved in the failure detection task areprimarily related to perceptual and decisionmakingmechanisms, the added resources required by tracking (asopposed to monitoring) concern more directly the responsemechanisms. To the extent that perceptual/decisionmakingprocesses and response processes draw from different struc-tural pools of processing resources that are not mutuallyavailable [24], [25], [31], [32], it is not expected that largeinterference would be evident between tracking anddetection.

This argument, relating to the lack ofinterference betweenstructurally different processes can also account for thedifferent effects upon detection and tracking performance ofa) introduction of the critical task (Fig. 1), and b) increases in

its difficulty (Fig. 2). Critical task introduction adds both a

new display element (demand for perceptual resources), as

well as new response demands required by the left controlmanipulation (demand for response resources). Thus theexpected decrease in both the concurrent tracking perfor-mance (Fig. 3) and in detection performance (Fig. 1) can bepredicted as their respective processing resources are bothdepleted by the added perceptual and response demands.However, the increase in i produces primarily an addeddemand on the availability of resources at the responsestages of processing, since the perceptual nature of thecritical task is little altered (critical task error changed littlewith 4), but a greater motor involvement is required. To theextent that the increase in l thus depletes response-relatedresources, but not perceptual ones, greater deterioration willbe evident in the response loading task (tracking) than inthe perceptual one (detection).

SUMMARY AND CONCLUSIONSThe major results can be briefly summarized as follows.1) For the conditions investigated in the present study,

detection of step increases in system order when the opera-

tor remains in the control loop (MA mode) is considerablyfaster and only slightly less accurate than when he isremoved (AU mode). This finding of MA superiority is

consistent with the earlier conclusions of Young [2].2) The extent of this superiority did not diminish as the

critical task was added or as its difficulty was increased byincreasing the subcritical value of A; an interaction betweenparticipatory mode and workload was not obtained.

3) The effect of adding the critical task was to reducedetection performance in both modes, but performance was

little altered with increasing A.4) Within the framework of the model presented, con-

verging evidence from multiple regression, ensemble averag-

ing, and latency distribution data was presented suggestingthat the cause of MA superiority was the added propriocep-

tive information, resulting from control adaptation andavailable for the first few seconds follwing the failure. Thisinformation, when coupled with the displayed visual infor-mation, allowed a rapid initial aggregation ofevidence in theMA mode, yielding short latency detections. However, theavailability of the proprioceptive adaptation informationwas short-lived, due perhaps to the transient memory forthe proprioceptive standard. Once gone, the now-adaptedvisual error signal continued to provide evidence of lesserstrength, accumulated at a slower rate than the nonadaptingAU condition. Therefore, the mean and modal latencies ofAU detections were longer, but overall detection wasslightly more accurate.There is a second interpretation that the greater AU

latency could be attributed to a greater loss of vigilance orwandering of focal attention from the display in the lessdemanding AU condition. While eye fixations were notrecorded, and so this hypothesis cannot be discounted, fourfactors argue against it. 1) Trial duration (2' minutes) andinterstimulus interval (approximately 25 seconds) were bothconsiderably shorter than those conventionally employed invigilance research. 2) The classic "vigilance decrement" [33](drop in performance across a trial) was not observed in thedata. 3) Detection accuracy was slightly enhanced in theAUcondition, a result contrary to that which would bepredicted by the loss of vigilance. 4) A subsequent study byKessel and Wickens [34] in which explicit payoffs wereprovided for detected failures (presumably attenuating anyeffect of vigilance) did not alter the magnitude of the latencydifference in detection.

5) The role of task workload in affecting detection perfor-mance was seemingly only evident as perceptual load wasincreased (adding the requirement to process the criticaltask display) and not as additional demands were placedupon the response systems, either through tracking theprimary task or through the greater resource demands of theincreased loading task difficulty.

6) Concerning the role of a more stable internal model asa cause of MA superiority, it can be argued that this factorprobably played a relatively minimal role in influencing thepresent results. This is because the repeated measures designallowed the same subjects to participate alternately on AUand MA trials. Thus the internal model constructed duringMA trials, if superior, was presumably also available on AUtrials. On this basis, it may be hypothesized that even greaterMA superiority might be obtained if participatory modewere manipulated as a between subjects variable. Con-versely, if the operators employed different models in thetwo conditions, then, given the predicted inferiority of theAU model, this difference might have contributed tothe obtained results.

7) The difference between the findings of Young and ofEphrath and Curry regardingMA versus AU superiority areconsistent with the current results. The MA condition inYoung's experiment and in the current one were in manyrespects similar. A step change in system order was imposedfollowing which control adaptation was required. Thisentailed a change in general response characteristics (opera-

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNE[I(Sc, "0L. SMC-9, No) I) -NtSRN' 197i

tor describing function) and therefore made proprioceptiveinformation available. Conversely, the failure employed by

Ephrath, a gradual displacement or bias of the flight path, is

one for which no fundamental change in the operator'stransfer function was required to adapt. Therefore, pro-

prioceptive channels probably conveyed little if any infor-mation relating to the occurrence of a failure. Furthermore,the ramp characteristic of Ephrath and Curry's failurewould have eliminated the presence ofany initial transient in

control activity. Since it is argued here that proprioceptive

information is only available (or utilized) early in thepost-failure interval, the absence of such a transient providesanother argument that the proprioceptive channel was notavailable in the Ephrath and Curry study, thereby eliminat-ing one potential source of MA superiority.

Also, it should be noted that while the autopilot employedYoung was adaptive, thereby attenuating the strength of

the visual signal, the autopilot in the present investigationwas not adaptive. Despite the advantage to detection thatthis provided, AU detection was still inferior.

8) Finally, some mention should be made concerning thepresence of individual differences. To some extent these are

inevitable, particularly in a task configuration as complex as

the current one, requiring dual task performance in the AUmode and triple task performance in the MA. Given thesubject's flexibility to allocate resources differentially to thetwo or three tasks, as well ashis ability to adopt various

criteria on the speed-accuracy detection bias, it is perhapssomewhat surprising that the individual subject data in Figs.1 and 2 are as consistent as they are.

ACKNOWLEDGMENT

The authors acknowledge the invaluable programming

assistance of R. Hunt in the conduct of this research. Dr. A.Fregly was the scientific monitor of the contract.

REFERENCES[1] T. B. Sheridan, "Preview of models of the human monitor/

supervisor," in Monitoring Behavior and Supervisory Control, T. B.Sheridan and G. Johannsen, Eds. New York: Plenum, 1976.

[2] L. R. Young, "on adaptive manual control," IEEE Trans. Mani-Mach. Syst., vol. MMS-10, pp. 292-331, 1969.

[3] D. Vreuls et al., 'Pilot failure detection performance with three levelsof fault warning information," Bunker-Ramo Corp., rep. no.

SRDS-RD-68-9, 1968.[4] A. R. Ephrath and R. E. Curry, 'Detection by pilots of system failures

during instrument landings," IEEE Trans. Syst., Man, Cyberni., vol.SMC-7, no. 12, pp. 841-848, 1977.

[5] W. Veldenhysen and H. G. Stassen. "The internal model: What doesit mean in human control," in Monitorinig Behavior and SupervisoryControl, T. B. Sheridan and 6. Johannsen, Eds. New York: Plenum,1976.

[6] R. W. Pew, "Human perceptual-motor performance," in Human In-formation Processing Tutorials ii Performance and Cogniition, B. H.Kantowitz, Ed. New York: Wiley, 1974.

[7] R. E. Curry and E. G. Gai, "Detection ot random process failures byhuman monitors," in Moniitorinig Beaavior and Supervisory Conitrol,T. B. Sheridan and G. Johannsen, Eds. New York: Plenum, 1976.

[8] D. C. Miller and J. J. Elkind, "The adaptive response of the humanicontroller to sudden changes in controlled process dynamics," IEEE

Tranis. Hum. Factors Electron., vol. HFE-8, pp. 218-223, 1967

[9] D. L. Kleinman, S. Baron, and W. H. Levison, "An optimal control

model of human response, Part 1: Theory and validation." 4utoma-tica, vol. 6, pp. 357--369, 1970.

[10] G. Stelmach, Motor Conitrol Issuies & Trenids, New York: Academlic1976.

[11] S. F'idell, "Sensory function in multidimensional signal detection."J.4coust. Soc. 4m., vol. 47, pp. 1009--1015, 1970.

[12] R. E. Curry and A. R. Ephrath. "Monitoring and control ot unl-reliable systems," in Monitoring Behavior and Supervisory Contrtol. T.B. Sheridan and G. Johannsen, Eds. New York: Plenum, 1976.[13] J. Adams, D. Gopher, andG. Lintern, "Effects of visual and pro

prioceptive feedback on motor learning.".l. Mot. Be/i., vol9* ppl.11-22, 1977.

[14] T. C. Jordan, "Characteristics of visual and proprioceptive response

times in thelearning of a motor skill." Quart. J. Exp. Psiych.. vol. 24,pp. 536-543, 1972.

[15] R. M. Klein and M.I. Posner, 'Attention tovusual and kinestheticcomponents of skills," Brain Res.. vol. 71, pp. 401--411. 1974.[16] J. D. McDonnell, "A preliminary study of human operator belhaviorfollowing stepchanges inthe controlled element." IEEE7'rliis.s Hiumt.Factors Electron., vol. HFE-7, pp. 125- 129, 1966.

[17]S. W. Keele.4ttention antd Humani Pertorman e. California'(ioodyear, 1973

[18] B. Kerr, "Processing demands during mental operationis."-lemornoatnd Cogtitioni, vol. 1, pp. 401-412, 1973.

[19] G. Johannsen, C. Pfendler, and W. Stein, "Hum ani performance and

workload in simulated landing-approaches with autopilot-failunies,"in Xlonitoring Belia(orr and Superiisory ControlT. B. Sheridan andG. Johannsen, Eds. New York: Plenum.1976.

[20] R. Jex, J.D. McDonnell, and A. V. Phatak, "A 'critical' tracking task

for manual control research." IEEE TranisHumsw1FLctorrs Electron.,vol. HFp7. pp. 138 144, 1966

[21] D. M. Green and J. A.Swets, Signial Detection Theoryv anid Psyclioph-vsics. New York: Wiley, 1966.

[22] C. S. Watson and T. L. Nichols. "Detectability of auditory signals

presented without defined observation intervals,"J. I1'oustic Soc,4mer.volv 59,pp. 655 668, 19766

[23] D. McNicol-.4 Primer of Siqnial Detection Theori% London: Allen aridIrwin, 1972.

[24] B. H. Kantowitz and S. L. Knight, "Testingtappiigtim-nesarinig: 11,Auditory secondary task,"clcta P.s yhologica,vol. 40. pp. 343 362,19761

[25] C. D. Wickens, J. Isreal, and E.Donchin. "The event related cortical

potential as an index of task workload." Proc. I4nn. Meetiiq of theHuman Factors Soc.. San Francisco, CA, Oct.. 1977

[26] R. Pachella. "The interpretation of reaction time in informationprocessing research," in HumaniItonformiation Processing, B.Kantow-it7, Ed. Hillsdale, NJ: Lawrence Erlbaum,1974.

[27] R. W. Pew, "The speed accuracy operating chiaiacteristic" ItoPstycliologica, vol 30, pp. 16 26. 1969

[28] D. T. McRuer and H. R. Jex, "A review of quasilinear pilot mnodels."IEEE Tranis. Hum. Factors Electron., vol.HFE-6. pp. 62 65,1965.

[29] D. H. Weir and A. V. Phatak, "Model of human operator response to

step transitions in controlled element dynamics." Seconid .Inn. Confon ManuualConttrol, NASASP-128,pp.65--83, 1966.

[30] J. Lappin and K.Disc, , "The latency operating characteristic: IEffects of stimulus probability." J. Ex per. Psych., vol. 92, pp. 419 427.1972.

[31] A. S. Sanders, in .lental W/orkload. Thieory and Iea.surement. N.Moray, Ed. New York: Plenum. 1978 (in press).

[32] C. D. Wickens, in Xlental Workload: Thieor'y' andd Xleasuremtent, N.Moray, Ed. New York: Plenum, 1978 (in press')

[33] J. Mackworth, V igilance and Habituationi. Middlesex.Enggland:Penguin, 1969.[34] C. Kessel and C. Wickens. "ThTe internal model: A study of tte rela-

live contribution of proprioception and vision to failure detectio*."14th Annu. Conf. Manual Control, I os Angeles. CA, 1978.

34