Effects of Preview on Human Control Behavior in Tracking Tasks With Various Controlled Elements

Abstract

This paper investigates how humans use a previewed target trajectory for control in tracking tasks with various controlled element dynamics. The human's hypothesized "near" and "far" control mechanisms are first analyzed offline in simulations with a quasi-linear model. Second, human control behavior is quantified by fitting the same model to measurements from a human-in-the-loop experiment, where subjects tracked identical target trajectories with a pursuit and a preview display, each with gain, single-, and double-integrator controlled element dynamics. Results show that target-tracking performance improves with preview, primarily due to the far-viewpoint response, which allows humans to cancel their own and the controlled element's lags, without additional control activity. The near-viewpoint response yields better target tracking at higher frequencies, but requires substantially more control activity. The control-theoretic approach adopted in this paper provides unique quantitative insights into human use of preview, which can help to explain human behavior observed in other preview control tasks, like driving.

Full text

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 1
Effects of Preview on Human Control Behavior
in Tracking Tasks with Various Controlled Elements
Kasper van der El, Student Member, IEEE, Daan M. Pool, Member, IEEE,
Marinus (Ren´
e) M. van Paassen, Senior Member, IEEE, and Max Mulder
Abstract—This paper investigates how humans use a previewed
target trajectory for control in tracking tasks with various con-
trolled element dynamics. The human’s hypothesized “near” and
“far” control mechanisms are first analyzed offline in simulations
with a quasi-linear model. Second, human control behavior is
quantified by fitting the same model to measurements from a
human-in-the-loop experiment, where subjects tracked identical
target trajectories with a pursuit and a preview display, each with
gain, single-, and double-integrator controlled element dynamics.
Results show that target-tracking performance improves with
preview, primarily due to the far-viewpoint response, which
allows humans to cancel their own and the controlled element’s
lags, without additional control activity. The near-viewpoint
response yields better target tracking at higher frequencies, but
requires substantially more control activity. The control-theoretic
approach adopted in this paper provides unique quantitative
insights into human use of preview, which can explain human
behavior observed in other preview control tasks, like driving.
Index Terms—Human control models, man-machine systems,
manual control, parameter estimation, preview control
I. INT ROD UC TI ON
HUMANS are highly effective adaptive controllers [1].
The seminal work of McRuer and his coworkers [2]
shows that Human Controllers (HC) systematically adapt their
control response to the dynamics of the Controlled Element
(CE), the display type, and the characteristics of the target
signal to be tracked. The HC’s adaptation mechanisms are rela-
tively well-understood in simple error-compensation tasks [3];
however, few practical control tasks are purely compensatory.
Instead, preview information of the target trajectory is often
visible, commanding the HC were to steer to in the near future.
Driving a car over a road is perhaps the best known exam-
ple [4]–[6], but most vehicle control tasks involve preview, as
well as many everyday motor control tasks [7], [8].
It has been shown that preview information helps HCs to
improve task performance, compared to zero-preview (pursuit)
tasks [5], [6], [9]–[11]. In tracking tasks, the amount of
preview needed for maximum performance depends, at least,
on the CE dynamics, and increases from about 0.5 to 1 s from
position to acceleration control tasks [9]–[11]. To extrapolate
these results to yet untested preview control tasks, many
cybernetic models have been proposed (e.g., see [5], [6],
[10]–[13]). Although several models accurately replicate the
human’s control outputs, they are unsuitable to systematically
study HC adaptation, because the proposed model inputs and
The authors are with the Control and Simulation section, Faculty of
Aerospace Engineering, Delft University of Technology, 2629 HS Delft, The
Netherlands. Corresponding author: K.vanderEl@tudelft.nl
multiloop control dynamics were never shown to resemble
those of the actual HC with objective measurements.
Recently, we measured the HC’s control dynamics in pre-
view tracking tasks using a multiloop frequency-domain sys-
tem identification technique [14]. Based on this, we extended
McRuer et al.’s [2], [3] quasi-linear model for compensatory
tracking tasks with two distinct responses that are based
on a “near” and a “far” viewpoint on the previewed target
ahead. The model’s physically interpretable parameters, like
the viewpoints’ locations, can be explicitly estimated from
measurement data. Thereby, this model may finally allow
for quantifying HC control adaptation in preview tracking
tasks, similar as established previously for compensatory
tracking [1]–[3]. Unfortunately, the model’s near- and far-
viewpoint responses are still poorly understood: while HCs
always apply a far-viewpoint response, the presence of a near-
viewpoint response appears to depend strongly on the tested
subject and CE dynamics [14]. It is unclear when and why
it is beneficial to respond to either one or two points on the
previewed target ahead.
The goal of this paper is to explain how HCs use preview for
control in manual tracking tasks with various CE dynamics.
We first investigate the roles of the near- and far-viewpoint
responses through offline simulations with the new preview
model from [14], with gain, single-, and double-integrator
CE dynamics. Second, we verify these offline predictions
with measurements from a human-in-the-loop experiment, in
which subjects performed a combined target-tracking and
disturbance-rejection task with these same CEs, both in tasks
with zero preview (i.e., pursuit) and 1 full second of preview.
These experimental data were also used to derive the preview
model in [14]; however, in this paper, we present a variety
of new measures. Effects of preview are quantified with
measures for tracking performance and control activity, and
with estimates of input-to-error and open-loop dynamics. The
HC’s underlying control behavior is investigated with non-
parametric estimates of their multiloop response dynamics, and
with estimates of the new preview model’s parameters [14].
This paper is structured as follows. In Section II, we
summarize important aspects of HC behavior in preview
tracking tasks, including the HC model from [14]. Offline
model analyses are presented in Section III. The performed
experiment and data analysis procedures are presented in
Section IV, followed by the experimental results in Section V.
We discuss these results and present our conclusions in the
final two sections of this paper.
K. van der El, D. M. Pool, M. M. van Paassen, and M. Mulder,
“Effects of preview on human control behavior in tracking tasks with various controlled elements,”
IEEE Trans. on Cybernetics, 2017.
DOI: 10.1109/TCYB.2017.2686335

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 2
II. BACKGROUND
A. The Control Task
The general layout of a combined target-tracking and
disturbance-rejection control task is illustrated in Fig. 1. In
these tasks, HCs are to minimize the tracking error:
e(t) = ft(t)−x(t),(1)
which is the difference between the current values of the target
signal ft(t)and the CE output x(t). HCs generate control
inputs u(t)to minimize this tracking error. At the same time,
the CE is perturbed by disturbance signal fd(t), for which the
HC must also compensate. In pursuit tasks, only the current
target at time tis presented on the display, together with the
CE output (see Fig. 2a). In preview tasks, an additional stretch
of the future target ft([t,t+
τ
p]) is visible, up to preview time
τ
ps ahead (see Fig. 2b).
B. Classical Approach and Results
HCs can adopt a multi-channel control organization in
pursuit and preview tracking tasks, initiating an independent
response to the target, the CE output, and the error, and in
preview tasks also to the target ahead [2], [15]–[17]. Because
explicit identification of all individual response dynamics is
impossible [15], [17], HC behavior in these tasks has been
traditionally analyzed by identifying lumped response dynam-
ics [9], [11], [15], [17]. Ito & Ito [11], for example, measured
the closed-loop dynamics from the target to the CE output:
Hcl,t(j
ω
) = X(j
ω
)
F
t(j
ω
),(2)
with Xand F
tthe Fourier transforms of the respective signals.
Perfect target-tracking is achieved when X(j
ω
)=F
t(j
ω
), or
equivalently, when |Hcl,t(j
ω
)|=1 and 6Hcl,t(j
ω
)=0 deg. Ito
& Ito’s results (partly reproduced in Fig. 3) reveal that preview
yields improved closed-loop characteristics, compared to the
e(t)u(t)x(t)
fd(t)
ft([t,t+τp])
x(t)
human
controller side-stick controlled
element
display
(pursuit/
preview)
ft
Fig. 1. The HC in a target-tracking and disturbance-rejection task.
re f erence target
ft(t)
e(t)
x(t)
controll ed
element
out put
(a)
τpft(t+τ)
τ
preview
t
(b)
Fig. 2. Layout of the pursuit (a) and preview (b) displays.
10-1 100101
10-1
100
ω
, rad/s
|Hcl,t|, -
10-1 100101
-360
-180
0
pursuit,
τ
p=0 s
preview,
τ
p=1 s
perfect tracking
ω
, rad/s
6Hcl,t, deg
Fig. 3. Closed-loop dynamics in a double-integrator task with and without
preview, average of two subjects (reproduced from [11]).
pursuit task, as the phase of Hcl,t(j
ω
)is closer to zero. The
closed-loop magnitude does not show a clear improvement.
Preview thus primarily helps HCs to better synchronize the CE
output with the target. In tasks with lower-order CE dynamics
(e.g., a gain), HCs extend the region where the closed-loop
phase approximates zero to higher frequencies [11]. Unfor-
tunately, the (lumped) closed-loop dynamics obscure exactly
how HCs use the available preview information, and also how
they adapt their control response to the CE dynamics.
C. Human Controller Model for Preview Tracking
Recently, we proposed a new model for pursuit and preview
tracking tasks that separates the HC’s responses to the differ-
ent input signals [14]. Thereby, this model provides deeper
insights in the human’s underlying control mechanisms.
1) The Model for Pursuit Tracking: The HC model for
pursuit tasks (see Fig. 4a) extends McRuer et al.’s simplified
precision model for compensatory tracking [3]. The model is
also quasi-linear, which means that linear describing functions
account for the linear portion of the HC’s response. Possible
nonlinear and time-varying behavior are not explicitly mod-
eled, nor are perception and motor noise; these are injected
together as filtered white noise through the remnant n(t).
The pursuit model involves a response to an error e⋆(t), with
response dynamics Hoe⋆(j
ω
)that are equal as in McRuer’s
simplified precision model [3], [14]:
Hoe⋆(j
ω
) = Ke⋆1+TL,e⋆j
ω
1+Tl,e⋆j
ω
.(3)
Ke⋆is the error response gain and TL,e⋆and Tl,e⋆are the lead
and lag time constants, respectively. Similar as in compen-
satory tracking, HCs adapt to the CE dynamics by generating
lead or lag in Hoe⋆(j
ω
), to establish a fair stretch of integrator-
like dynamics around the open-loop crossover frequency (
ω
c):
|Hoe⋆Hce|≈
ω
c/j
ω
[14], [15].
The error e⋆(t), a signal internal to the HC, is defined as the
difference between the filtered target f⋆
tand the CE output:
E⋆(j
ω
) = F⋆
t(j
ω
)−X(j
ω
) = Hof(j
ω
)F
t(j
ω
)−X(j
ω
).(4)
In pursuit tasks, Hof(j
ω
)was modeled as a simple gain,
Hof(j
ω
)=Kf[14]. When Kf=1, (4) shows that e⋆(t)=e(t),
hence that HCs respond to the true error and that they
effectively exhibit a single-channel “compensatory” control
organization [18]. A non-unity value of Kfimplies a “pursuit”
control organization [18], or the presence of a feedforward

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 3
e⋆(t)u(t)x(t)
n(t)
human controller fd(t)
ft(t)
Hnmse−τvjω
HofHoe⋆
+
+
+
+
+
−
Hce
f⋆
t(t)
(a)
e⋆(t)
ft,n(t)
u(t)x(t)
n(t)
human controller
fd(t)
ft,f(t)
Hnmse−τvjω
Hof
Hon
Hoe⋆
+
+
+
+
+
+
+
−
Hce
f⋆
t,f(t)
(b)
Fig. 4. Control diagrams of the HC model for pursuit (a) and preview (b)
tracking tasks [14].
response. Higher values of Kfindicate a more aggressive
response to the target, while Kf=0 means that the HC com-
pletely ignores the target and focuses only on disturbance
rejection. Single-subject data showed that Kf<1 for an (un-
stable) double integrator CE, Kf≈1 for an integrator CE, and
Kf>1 for a (stable) gain CE [14], which suggests that Kf
reflects an important control-adaptation mechanism.
The model also incorporates the HC’s most dominant physi-
cal limitations. Visual response delay
τ
vcombines perceptual,
cognitive and transport delays, and Hnms(j
ω
)represents the
combined neuromuscular system (NMS) and side-stick dy-
namics:
Hnms(j
ω
) =
ω
2
nms
(j
ω
)2+2
ζ
nms
ω
nms j
ω
+
ω
2
nms
,(5)
with
ω
nms and
ζ
nms the natural frequency and damping ratio.
2) The Model for Preview Tracking: Fig. 4b shows the HC
model for preview tasks, which extends the pursuit model.
Two responses, each initiated with respect to a different view-
point, can capture the HC’s response to the entire previewed
target [14]. A far viewpoint ft,f(t)feeds the “pursuit” control-
loop, while an additive, parallel feedforward channel describes
the HC’s response to a near viewpoint ft,n(t). The near- and
far-viewpoints are located
τ
nand
τ
fs ahead on the previewed
target:
ft,n(t) = ft(t+
τ
n),ft,f(t) = ft(t+
τ
f).(6)
As the HC can select which points to respond to, based on the
task specifics, both
τ
nand
τ
fare free model parameters. Note
that these viewpoints do not necessarily correspond to the two
levels, or points, used in many driver models (e.g., [19], [20]).
In preview tracking tasks, HCs were found to smooth the
target in the far viewpoint, so Hof(j
ω
)includes a low-pass
filter [14]:
Hof(j
ω
) = Kf
1
1+Tl,fj
ω
.(7)
The far-viewpoint response thus only describes low-frequency
target-tracking behavior, with the reciprocal of the time con-
stant Tl,fas cut-off frequency. The HC’s response to higher
frequencies in the target signal was modeled as an open-loop
response Hon(j
ω
)with respect to the near viewpoint [14]:
Hon(j
ω
) = Kn
j
ω
1+Tl,nj
ω
,(8)
with gain Knand high-pass filter time-constant Tl,n. The
limited data provided in [14] suggests that not all subjects
apply a near-viewpoint response in tasks with single- and
double-integrator CE dynamics.
III. OFFLI NE MO DE L ANALYSIS
The exact roles of the near- and far-viewpoint responses
are not yet fully understood. To gain more insight, we math-
ematically derive the HC dynamics that result in “perfect”
target-tracking, and we investigate the contributions of both
responses with model simulations.
A. Perfect Target-Tracking
The introduced HC model (Fig. 4) can be restructured into
the mathematically equivalent two-channel model of Fig. 5
(see [14]). Here, the HC is modeled to respond to the target and
the CE output, with lumped dynamics Hot(j
ω
)and Hox(j
ω
):
Hot=HofHoe⋆e
τ
fj
ω
+Hone
τ
nj
ω
Hnmse−
τ
vj
ω
,(9)
Hox=Hoe⋆Hnms e−
τ
vj
ω
.(10)
In (9) and (10) the dependency on j
ω
is left out for better
readability. Using Fig. 5, the target closed-loop can be written
as
Hcl,t(j
ω
) = X(j
ω
)
F
t(j
ω
)=Hot(j
ω
)Hce(j
ω
)
1+Hox(j
ω
)Hce(j
ω
).(11)
Substituting X(j
ω
)/F
t(j
ω
)=1 (i.e., perfect target-tracking),
and solving for Hot(j
ω
), yields the perfect target-tracking
dynamics HP
ot(j
ω
):
HP
ot(j
ω
) = Hox(j
ω
) + 1
Hce(j
ω
).(12)
Because the form of the response function Hox(j
ω
)is identical
in tasks with and without preview for a given CE [14],
the form of HP
ot(j
ω
)is also fixed. For example, Hox(j
ω
)is
approximately a gain for integrator CE dynamics. 1/Hce(j
ω
)
is then a pure differentiator, which has a negligible magnitude
at low frequencies, but a much higher magnitude than Hox(j
ω
)
at high frequencies. HP
ot(j
ω
)is thus dictated by Hox(j
ω
)at
low frequencies and by 1/Hce(j
ω
)at high frequencies. The
modeled HC target response in (9) has a similar form; for
integrator CE dynamics, it is dictated by gain KfKe⋆at low
frequencies and by differentiator Knj
ω
at higher frequencies.
This suggests that HCs attempt to approach perfect target-
tracking when preview is available.
ft(t)
x(t)
Hce
fd(t)
u(t)
n(t)
+
−+
+
+
+
Hot
Hoxhuman
controller
Fig. 5. Two-channel control diagram of the HC.

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 4
B. Model Simulations
Two key aspects of the model are essential for the difference
between pursuit and preview tasks: 1) the point on the target
ahead that is the input to the HC’s “pursuit” response (charac-
terized by
τ
f), and 2) the presence and strength of the additive
open-loop near-viewpoint response (characterized mostly by
Kn, but to a lesser extent also by the other model parameters).
Next, we investigate these two aspects for gain, single-, and
double-integrator CE dynamics.
1) Settings: For tasks with 0 and 1 s of preview, esti-
mated model parameters (single-subject data) are reproduced
from [14] in Table I; these are used as baseline in the simula-
tions. The used target (
σ
ft=0.5 inch) and disturbance (
σ
fd=0.2
inch) signals are each the sum of 20 sines, with a square
amplitude spectrum (1.5 rad/s bandwidth), augmented with
a high-frequency shelf where the amplitudes are attenuated
(see [14] for details). No remnant is included.
2) Analysis of the Far-Viewpoint Location: We step-wise
increase the value of
τ
ffrom 0 s (i.e., pursuit tracking) to
1.5 s, while keeping all other parameters fixed at the pursuit
settings in Table I. Fig. 6a shows that the variance of the
tracking error reduces substantially when
τ
fincreases, for all
CE dynamics. Doing so, the target response exhibits phase lead
that compensates for the CE’s inherent lag, and the HC’s NMS
lag and visual response delay. The phase becomes markedly
closer to HP
ot(j
ω
), especially at mid-frequencies, as shown for
integrator CE dynamics in Fig. 6e. Responding to the target
ahead requires no additional control activity (constant
σ
2
uin
Fig. 6a), because a pure delay like
τ
fonly affects the phase
of the target response (all |Hot(j
ω
)|lines overlap in Fig. 6c).
Fig. 6a also shows that it is beneficial to respond to the target
farther ahead for higher-order CEs, to compensate for its larger
inherent lag.
3) Analysis of the Near-Viewpoint Response: We step-wise
increase the value of Knfrom 0 and 0.6, keeping all other
parameters fixed at the preview settings in Table I. Fig. 6b
shows that only a small performance improvement is possible
by increasing Kn, which comes at the cost of a substantially
TABLE I
TES TED C ON DIT IO NS AN D MO DEL PA RAM ET ERS (S ING LE -SUB JE CT DATA),
ADA PTE D FRO M [14].
Hce Kce Kce/s Kce/s2
Kce, - 0.8 1.5 5
τ
p, s 0 1 0 1 0 1
abbreviation GN0 GN1 SI0 SI1 DI0 DI1
Ke⋆, - 3.85 6.62 1.43 1.11 0.14 0.14
Tl,e⋆, s 2.06 2.39 - - - -
TL,e⋆, s - - - - 2.54 2.22
τ
v, s 0.18 0.16 0.23 0.18 0.28 0.31
ω
nms, rad/s 17.9 18.0 11.2 10.2 6.15 5.33
ζ
nms, - 0.18 0.37 0.30 0.26 0.67 0.50
Kn, - - 0.06 - 0.18 - 0.32
τ
n, s - 0.08 - 0.34 - 0.00
Tl,n, s - 0.06 - 0.04 - 5.89
Kf, - 1.21 1.11 0.95 1.12 0.54 0.63
τ
f, s - 0.55 - 0.70 - 0.99
Tl,f, s - 0.26 - 0.38 - 0.59
higher control activity. For some subjects no near-viewpoint
response was found in [14]; possibly, these subjects aimed for
lower control activity, instead of slightly better performance.
The Bode plots (Fig. 6 d and f) show that an additional
near-viewpoint response mainly affects the high frequencies of
Hot(j
ω
), which resembles HP
ot(j
ω
)better in both magnitude
and phase if Knis non-zero. In particular, the characteristic
increasing phase lead that results from responding to a far
viewpoint (due to negative delay
τ
f) disappears, even with
low values of Kn.
4) Analysis of Time-Traces: The simulated CE output is
calculated with (11) for both the pursuit and preview pa-
rameters in Table I, with the disturbance set to zero. Fig. 7
shows that the CE output follows the target signal much better
with preview, lagging less behind, which is consistent with
Fig. 6. Still, the fast oscillations, or high frequencies, are
not completely reproduced; the CE output often remains on
the inside of the target signal “corners”, reflecting corner-
cutting behavior. This corresponds well with |Hot(j
ω
)|at high
frequencies (Fig. 6d), which is smaller than than required
for perfect target-tracking when Knis small. With double
integrator CE dynamics the target’s high frequencies are hardly
0 0.5 1 1.5
0.0
0.1
0.2
0.3
0.4
0.5
e, GN
u, GN
e, SI
u, SI
e, DI
u, DI
∆
σ
2
e
σ
2, inch
τ
f, s
τ
opt
f
pursuit
(a)
0 0.2 0.4 0.6
0.0
0.1
0.2
0.3
0.4
0.5
σ
2, inch
Kn, s
Kopt
n
(b)
10-1 100101
10-2
10-1
100
101
ω
, rad/s
1/Hce
Hox
HP
ot
|Hot|, -
Hot(
τ
f)
(c)
10-1 100101
10-2
10-1
100
101
ω
, rad/s
1/Hce
Hox
HP
ot
Kn=0.6
Kn=0
|Hot|, -
Hot(Kn)
(d)
100
-360
-180
0
180
360
ω
, rad/s
τ
f=0
τ
f=1.5
6Hot, deg
(e)
10-1 100101
-360
-180
0
180
360
ω
, rad/s
Kn=0
Kn=0.6
6Hot, deg
(f)
Fig. 6. Simulated effects of
τ
f(a), (c), and (e), and Kn(b), (d), and (f); Bode
plots (c-f) show only integrator CE dynamics results.

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 5
12 14 16
-0.5
0.0
0.5
1.0
t, s
ft/x, inch
condition: GN
(a)
12 14 16
-0.5
0.0
0.5
1.0
t, s
ft/x, inch
ft
x, pursuit
x, preview
condition: SI
(b)
12 14 16
-0.5
0.0
0.5
1.0
t, s
ft/x, inch
condition: DI
(c)
12 14 16
-1.0
-0.5
0.0
0.5
1.0
t, s
ft/x, inch
f⋆
t,f
xf
xn
condition: SI1
(d)
Fig. 7. Simulated time-traces of the CE output: the benefit of preview (a-c),
and the contributions of the near- and far-viewpoint responses (d).
tracked at all (Fig. 7c).
Fig. 7d shows the contributions of the near- and far-
viewpoint responses when tracking with preview and integrator
CE dynamics. The near-viewpoint response accounts for an
output (Xn=HonHnms e−
τ
vF
t,n) that is identical to the high-
frequency sinusoids in the target signal. The filtered far-
viewpoint (F⋆
t,f=HofF
t,f) lacks exactly these high frequencies;
tracking it results in an output (Xf, obtained from closed-
loop simulation with Kn,
τ
nand Tl,nall set to zero) that
approximates the target signal’s low frequencies.
IV. MET HO D
Next, the model simulations are verified with experimental
data. Details of the experiment and the data analysis proce-
dures are presented in this section.
A. The Experiment
Twelve subjects performed a combined target-tracking and
disturbance-rejection tasks. Two independent variables were
varied, the display and the CE dynamics. The display (see
Fig. 2) showed either 0 (i.e., pursuit) or 1 s of preview; the
CE had gain, integrator, or double integrator dynamics. All
subjects performed the full factorial of the two independent
variables in a randomized order. The six experimental condi-
tions are summarized in Table I; full details of the experimen-
tal settings, procedure, and apparatus are given in [14].
B. Data Analysis
1) Error and Control Output Variance: The variances of
the tracking error and the control output are used as measures
for the achieved tracking performance and the applied control
activity, respectively. The individual contributions due to the
target, disturbance, and HC remnant are estimated by integrat-
ing the error and control output auto spectral-density functions
only over the respective signal’s input frequencies [21].
2) Input-to-Error Dynamics: The target-to-error and
disturbance-to-error dynamics, Hft,e(j
ω
t)and Hfd,e(j
ω
d), re-
spectively, quantify the error amplification/attenuation, relative
to the respective input signal, in the frequency domain. Both
are estimated at the input signal’s frequencies,
ω
tor
ω
d, as
follows:
Hft,e(j
ω
t) = E(j
ω
t)
F
t(j
ω
t),Hfd,e(j
ω
d) = E(j
ω
d)
Fd(j
ω
d).(13)
3) Open-Loop Dynamics: In the frequency domain, per-
formance and stability are characterized by the open-loop
crossover frequency
ω
cand phase margin
φ
m, respectively.
In a combined target-tracking and disturbance-rejection task,
two open-loop dynamics can be formulated, Hol,t(j
ω
)and
Hol,d(j
ω
)[21]:
Hol,t(j
ω
t) = X(j
ω
t)
E(j
ω
t)
=Hot(j
ω
t)Hce(j
ω
t)
1+ [Hox(j
ω
t)−Hot(j
ω
t)]Hce(j
ω
t),(14)
Hol,d(j
ω
d) = −X(j
ω
d)−Fd(j
ω
d)
X(j
ω
d)
=Hce(j
ω
d)Hox(j
ω
d).(15)
Crossover occurs at the frequency
ω
cfor which |Hol(j
ω
)|=1,
the corresponding phase margin
φ
mis 180+6Hol(j
ω
c)deg.
4) Non-Parametric Multiloop System Identification: Non-
parametric estimates of Hot(j
ω
)and Hox(j
ω
)in Fig. 5
are used to objectively quantify the HC’s multiloop control
dynamics. Both responses can be estimated simultaneously
with a system identification method based on Fourier coeffi-
cients [14], [22], [23]. From Fig. 5 it follows that the modeled
control output is
U(j
ω
) = Hot(j
ω
)F
t(j
ω
)−Hox(j
ω
)X(j
ω
) + N(j
ω
).(16)
Two equations, needed to solve for the two unknown dy-
namics, are obtained by evaluating (16) both at the input
frequencies
ω
tof target signal, and by interpolating the signals
F
t,X, and Uin the frequency domain from the disturbance
frequencies
ω
dto these same
ω
t(yielding ˜
F
t,˜
X, and ˜
U).
Assuming zero remnant, it follows that
U(j
ω
t)
˜
U(j
ω
t)=F
t(j
ω
t)−X(j
ω
t)
˜
F
t(j
ω
t)−˜
X(j
ω
t)Hot(j
ω
t)
Hox(j
ω
t).(17)
Eq. (17) can be solved for Hot(j
ω
t)and Hox(j
ω
t). Similarly,
estimates can be obtained at the disturbance signal input
frequencies, by evaluating (17) at
ω
d, after interpolating from
ω
tto
ω
d.
5) Model Parameter Estimation: Estimates of the model
parameters are used to explicitly quantify human control
behavior, including the characteristics of the near- and far-
viewpoint responses. The model parameters are estimated by
minimizing a least-squares cost function J, which is based
on a frequency-domain error Eubetween the measured and
modeled control outputs Uand ˆ
U, respectively:
Eu(j
ω
|Θ) = U(j
ω
)−ˆ
U(j
ω
|Θ),(18)
J(Θ) =
Nl
∑
l=1

Eu(j
ω
l|Θ)


2.(19)

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 6
Nlis the number of measured frequencies below a chosen
cut-off frequency, here 25 rad/s. The five-run frequency-
domain average of the measured control output signals is
used to reduce effects of the remnant on the parameter
estimates. The modeled control output is obtained from (16)
with remnant Nset to zero. The parameter vector Θis
[Ke⋆Tl,e⋆TL,e⋆
τ
v
ω
nms
ζ
nms Kf
τ
fTl,fKn
τ
n]T. Because
the break frequency of the near-viewpoint high-pass filter was
generally well above measured frequency range in [14], Tl,n
is removed from the model here, such that (8) simplifies
to a pure differentiator. NMS natural frequencies above the
highest input frequency (about 15 rad/s) cannot be estimated
accurately, for subjects where this applies we fix
ω
nms at 15
rad/s. A Nelder-Mead simplex algorithm is used to minimize
J, constrained only to avoid solutions that contain negative
parameters. The best solution is selected from 100 randomly
initialized optimizations.
6) Data Processing: All non-parametric measures are cal-
culated per run, and then averaged over the five measurement
runs. Crossover frequencies and phase margins are calculated
from the fitted HC model, which allows for better estimates of
crossover frequencies outside the range of input frequencies.
A repeated-measures two-way ANOVA is applied to test for
significant differences in performance and control activity,
crossover frequency, and phase margin; results are compen-
sated with a conservative Greenhouse-Geisser correction when
the assumption of sphericity is violated. Errorbars on the
results in the next section represent 95% confidence intervals,
corrected for between-subject variability.
C. Hypotheses
Preview is information about the future target signal, so
we expect that it affects only the target-tracking, and not
the disturbance-rejection part of the task. This leads to the
following hypotheses:
I: Target-tracking performance improves with preview, in
accordance with [9]–[11] and our offline model predic-
tions; this will manifest in a lower error variance at the
target frequencies and higher target crossover frequencies
and phase margins;
II: Disturbance-rejection behavior is similar in pursuit and
preview conditions, resulting in similar control output
variances and Hox(j
ω
)dynamics, hence similar param-
eters Ke⋆,Tl,e⋆,TL,e⋆,
τ
v,
ω
nms, and
ζ
nms.
Based on our offline model analyses (Section III-B) we further
hypothesize that:
III: Subjects respond to the target ahead to improve perfor-
mance (characterized by
τ
nand
τ
f); furthermore, the two
viewpoints are farther ahead in conditions with higher-
order CE dynamics, to generate more compensating phase
lead for the CE’s larger inherent phase lag;
IV: Subjects initiate a weak near-viewpoint response, re-
flected by a small but non-zero value of Kn, to better
match the phase required for perfect target-tracking, with-
out substantially increasing control activity.
V. RESULTS
A. Tracking Performance and Control Activity
Fig. 8a shows that tracking performance is substantially bet-
ter (lower
σ
2
e) in conditions with preview, which corresponds
to results in [9]–[11]. Especially target-tracking performance
improves (gray part of the bars), but the slight performance
increase due to reduced HC remnant is also significant (see
Table II). Neither disturbance-rejection performance, nor con-
trol activity (Fig. 8b), are significantly different with preview.
Fig. 8 also shows that the performance improvement predicted
by the model simulations in Section III-B matches reasonably
well with the experimental results.
With higher-order CE dynamics, tracking performance is
substantially worse (Fig. 8a). However, this effect is smaller
when preview is available, especially at the target and remnant
frequencies (significant interaction effects). Increasing the
order of the CE dynamics markedly affects the control activ-
ity distribution: the target component decreases significantly,
while the remnant component increases significantly.
The estimated input-to-error dynamics are shown in
Fig. 9 for integrator CE dynamics. The characteristic error-
amplification peak, caused by the HC’s response time-
delay [3], is clearly present in disturbance rejection, both with
and without preview (indicated by |Hfd,e|>1 in Fig. 9b). In
target tracking (Fig. 9a) this peak is only visible in pursuit
0.0
0.1
0.2
0.3
0.4
0.5
GN0 GN1 SI0 SI1 DI0 DI1
σ
2
e, inch2
target
disturbance
remnant
∆
σ
2, pred.
τ
f
∆
σ
2, pred. Kn
∆
σ
2
e
(a)
0.0
0.2
0.4
0.6
GN0 GN1 SI0 SI1 DI0 DI1
σ
2
u, inch2
(b)
Fig. 8. Variances of the tracking error (a) and the control output (b).
TABLE II
ERRO R AN D CON TRO L OUT PU T ANOVA RE SULT S.1
error, econtrol output, u
df F sig. df F sig.
display (1,11) 127 ** (1,11) 1.31 -
σ
2dynamics (1.07,11.8) 213 ** (1.23,13.6) 8.4 **
disp.*dyn. (1.05,11.5) 13.5 ** (2,22) 0.91 -
display (1,11) 305 ** (1,11) 1.44 -
σ
2
tdynamics (1.16,12.8) 83.3 ** (2.22) 197 **
disp.*dyn. (1.11,12.3) 14.5 ** (2.22) 0.95 -
display (1,11) 0.22 - (1,11) 0.81 -
σ
2
ddynamics (1.01,11.1) 138 ** (1.08,11.9) 2.30 -
disp.*dyn. (1.03,11.3) 0.20 - (1.36,15.0) 0.79 -
display (1,11) 1.43 * (1,11) 1.34 -
σ
2
rdynamics (1.03,11.3) 135 ** (1.16,12.7) 14.2 **
disp.*dyn. (1.08,11.8) 7.98 * (2,22) 1.10 -
1Symbols **, *, and - indicate highly significant (p< .01), significant
(p< .05), and non-significant (p> .05) results, respectively.

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 7
10-1 100101
10-2
10-1
100
target
ω
, rad/s
|Hft,e|, -
(a)
10-1 100101
10-2
10-1
100
disturbance
difference
pursuit
preview
ω
, rad/s
|Hfd,e|, -
(b)
Fig. 9. Non-parametric estimates of the input-to-error dynamics for integrator
CE conditions, single-subject data.
tasks. With preview, |Hft,e|is always smaller than one, so the
error is attenuated all input frequencies. This is evidence that
preview enables HCs to compensate for their own response
delays.
B. Open-Loop Dynamics
In pursuit conditions, the measured target open-loop dy-
namics (Fig. 10, gray markers/line) resemble an integrator
with a time delay around crossover, in accordance with [15],
[24]. For double integrator CE dynamics subjects managed
to generate integrator magnitude characteristics in only a
minor region around crossover, due to the difficulty of this
condition. All disturbance open-loop dynamics (not shown)
have a similar shape, both in pursuit and preview conditions.
With the introduction of preview, the magnitude of the target
open-loop dynamics increases below the crossover frequency,
and then drops off with a slope larger than that of an integrator
10-1 100101
10-2
100
102GN, magnitude
ω
, rad/s
ω
c,t
ω
c,t
|Hol,t|, -
10-1 100101
-540
-360
-180
0
180 GN, phase
pursuit, non-par.
preview, non-par.
ω
, rad/s
φ
m,t
φ
m,t
6Hol,t, deg
10-1 100101
10-2
100
102SI, magnitude
ω
, rad/s
ω
c,t
ω
c,t
|Hol,t|, -
10-1 100101
-540
-360
-180
0
180 SI, phase
pursuit, model
preview, model
ω
, rad/s
φ
m,t
φ
m,t
6Hol,t, deg
10-1 100101
10-2
100
102DI, magnitude
ω
, rad/s
ω
c,t
ω
c,t
|Hol,t|, -
10-1 100101
-540
-360
-180
0
180 DI, phase
ω
, rad/s
φ
m,t
φ
m,t
6Hol,t, deg
Fig. 10. Target open-loop dynamics, single-subject data.
(black markers/line, Fig. 10); additionally, the characteristic
pure delay is not visible in the open-loop phase.
The target crossover frequency (Fig. 11a) and phase mar-
gin (Fig. 11c) are both higher in conditions with preview
(significant effect, Table III), pointing to improved target-
tracking performance and stability. The average target phase
margins are between the values predicted by the near- and far-
viewpoint model simulations (Section III-B), suggesting that a
combination of both responses is active (except in double inte-
grator tasks). Note that the measured crossover frequencies are
slightly lower than the idealized predictions. The disturbance
crossover frequency (Fig. 11b) and phase margin (Fig. 11d)
are similar in pursuit and preview conditions. Only for gain CE
dynamics the disturbance crossover frequency is slightly lower
with preview, yielding a significant display effect; however,
this crossover frequency was difficult to estimate, due to the
relatively low control activity at disturbance frequencies in
gain CE conditions (see Fig. 8b).
The measured crossover frequencies (except target tracking
with preview) are relatively low: they are in the region where
crossover regression occurs in compensatory tracking tasks
(0.8
ω
c<
ω
i[2], [25]), as illustrated in Fig. 11. Little is
0
4
8
12 target
GN SI DI
ω
c,t, rad/s
(a)
0
1
2
3
4disturbance
regression limit
GN SI DI
ω
c,d, rad/s
ω
i
(b)
0
100
200
300 target
GN SI DI
φ
m,t, deg
∆, pred.
τ
f
∆, pred. Kn
(c)
0
40
80
120 disturbance
pursuit
preview
GN SI DI
φ
m,d, deg
(d)
Fig. 11. Crossover frequencies (a,b) and phase margins (c,d).
TABLE III
CROS SOV ER FR EQU EN CY AN D PH ASE M AR GIN A NOVA RES ULTS .1
target disturbance
df F sig. df F sig.
display (1,11) 35.7 ** (1,11) 7.80 *
ω
cdynamics (1.20,13.2) 19.7 ** (1.27,14.0) 8.35 **
disp.*dyn. (1.23,13.5) 12.1 ** (2,22) 8.69 **
display (1,11) 37.7 ** (1,11) 0.41 -
φ
mdynamics (1.58,17.4) 19.9 ** (1.20,13.1) 316 **
disp.*dyn. (1.17,12.9) 8.32 ** (1.33,14.7) 1.40 -
1Symbols **, *, and - indicate highly significant (p< .01), significant
(p< .05), and non-significant (p> .05) results, respectively.

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 8
known of this phenomenon in pursuit and preview tasks. The
relative invariance of these low crossover frequencies with
CE dynamics was reported earlier in similar pursuit tracking
tasks [15], [26].
C. Human Multiloop Control Dynamics
Fig. 12 shows Bode plots of the estimated Hot(j
ω
)and
Hox(j
ω
). As shown before in [14], the model fits (solid lines)
coincide well with the non-parametric identification results
(markers). Note that a similar equalization is visible as in
compensatory tracking tasks [3]; both Hot(j
ω
)and Hox(j
ω
)
exhibit a -1, 0, and +1 mid-frequency magnitude slope for gain,
single-, and double-integrator CE dynamics, respectively.
The target response in pursuit conditions (as well as the
CE output response in all conditions), shows the characteristic
high-frequency phase roll-off caused by the HC’s response
delay and NMS lags. In the preview conditions such phase
lag is not present in Hot(j
ω
); instead, phase lead is generated,
similar as in the simulations in Section III-B. The resulting
phase characteristics resemble perfect target-tracking much
better, so subjects clearly apply control actions that cancel
most of the lag from their own response and the CE dynamics.
At higher frequencies and for higher-order CE dynamics the
perfect target-tracking phase is matched less well.
Fig. 12 also shows that the target response high-frequency
magnitude is lower than that required for perfect target-
tracking. This indicates corner-cutting behavior, and corre-
sponds to the model simulations with low values of the near-
viewpoint gain Kn(Section III-B).
D. Model Parameters
1) Internal-Error Response: In gain CE conditions, both
Ke⋆and Tl,e⋆increase slightly with preview (Fig. 13). As a
results, the total error-response dynamics have a higher mag-
nitude at the lowest frequencies, but remain similar over most
of the measured frequency range. Similarly, preview yields a
slightly higher Ke⋆and TL,e⋆in double integrator conditions,
which also points to a higher low-frequency magnitude. For
integrator CE tasks, Ke⋆is identical with and without preview.
2) Physical Limitations:
τ
v,
ω
nms, and
ζ
nms (Fig. 13) are
not systematically adapted when preview becomes available.
Only the NMS damping
ζ
nms appears to be slightly lower with
preview. Increasing the order of the CE dynamics yields more
pronounced effects: the visual-response delay
τ
vincreases,
while the NMS bandwidth (
ω
nms) decreases; such adaptations
have been measured before in [2], [15], [24], [26].
3) Far-Viewpoint Response: Fig. 14 shows the estimated
far-viewpoint parameters.
τ
fis larger for higher-order CE
dynamics, indicating that subjects respond to the target farther
ahead, to generate more compensating phase lead. For double
integrator CE dynamics,
τ
fis approximately at the limit
of the presented preview (1 s), suggesting that the tracking
performance in this condition may further improve with more
preview. The far-viewpoint filter time-constant Tl,fis also
larger for higher-order CEs, such that less of the target’s high
frequencies are tracked through the far-viewpoint response.
To compensate for the phase lag introduced by the low-pass
filter, the measured values of
τ
fare consistently higher than
predicted in Section III-B, where this low-pass filter was not
considered (i.e., Tl,f=0).
For gain and integrator CE dynamics, the target weighting
gain Kfis similar in pursuit and preview conditions. For dou-
ble integrator CE dynamics, Kfis much larger with preview,
indicating that subjects are responding more aggressive to the
target signal. The difficulty of the pursuit task with double
integrator CE dynamics likely forced subjects to prioritize
stabilizing the CE’s output, so less effort was put in target
tracking. This is consistent with the generally lower values
of Kfwith higher-order CEs, and also with the lower control
activity at the target frequencies (Fig. 8).
4) Near-Viewpoint Response: Fig. 15 shows the estimated
near-viewpoint parameters. Knis small but always non-zero,
suggesting that most subjects initiated a near-viewpoint re-
sponse; however, this does not correspond to the Bode plots
in Fig. 12. For example, for double integrator CE dynamics
the increasing high-frequency phase of Hot(j
ω
)suggests that
no near-viewpoint response is present, while Knis estimated at
0.05. For single integrator CE dynamics, the phase flattening
of Hot(j
ω
)at high frequencies does suggest that a near-
viewpoint response is initiated, while Knis estimated at 0.08. It
is thus difficult to determine whether a subject initiated a near-
viewpoint response, or not, merely from Kn. The adaptation
of Knto the CE dynamics is similar as predicted by the model
simulations (Section III-B), with the highest value of Knfound
in single integrator conditions. As the estimated values of Kn
are lower than predicted, it appears that subjects prioritize a
low control activity over enhanced performance.
τ
nis larger for higher-order CE dynamics, similar as
τ
f.
However, between-subject variations are large, especially for
double integrator CE dynamics. Likely, these variations (and
the outlier for Subject 5 with single integrator CE dynamics)
point to a negligible contribution of the near-viewpoint re-
sponse. Consequently, it is impossible to obtain a meaningful
estimate of
τ
n.
VI. DISCUSSION
In this paper, we explained how HCs use preview for
control in manual tracking tasks with various CE dynamics,
using both offline model analyses and experimental data.
The hypothesized performance improvement with preview, in
accordance with [9]–[11], was confirmed, predominantly in
target tracking (H.I). Offline model simulations predicted the
attained performance improvement remarkably well, especially
considering that no remnant was included, and no param-
eter interactions were investigated. As hypothesized (H.II),
disturbance-rejection behavior and performance were similar
with and without preview.
Fitting the model to the experimental data allowed us to
peek inside the black-box of human control, decomposing their
behavior into several characteristic responses and physically
interpretable parameters. Thereby, we confirmed that subjects
respond to the target farther ahead in tasks with higher-order
CE dynamics (as suggested before in [14]), to compensate for
the CE’s larger phase lag (confirming H.III). The adopted far-
viewpoint location was anticipated quite accurately with the

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 9
10-1 100101
10-2
10-1
100
101GN, target
pursuit, perfect
preview, perfect
ω
, rad/s
|Hot|, -
10-1 100101
-360
-180
0
180 GN, target
pursuit, model
preview, model
ω
, rad/s
6Hot, deg
10-1 100101
10-2
10-1
100
101GN, CE output
pursuit, non-par.
preview, non-par.
ω
, rad/s
|Hox|, -
10-1 100101
-360
-180
0
180 GN, CE output
ω
, rad/s
6Hox, deg
10-1 100101
10-2
10-1
100
101SI, target
ω
, rad/s
|Hot|, -
10-1 100101
-360
-180
0
180 SI, target
ω
, rad/s
6Hot, deg
10-1 100101
10-2
10-1
100
101SI, CE output
ω
, rad/s
|Hox|, -
10-1 100101
-360
-180
0
180 SI, CE output
ω
, rad/s
6Hox, deg
10-1 100101
10-2
10-1
100
101DI, target
ω
, rad/s
|Hot|, -
10-1 100101
-360
-180
0
180
360
DI, target
ω
, rad/s
6Hot, deg
10-1 100101
10-2
10-1
100
101DI, CE output
ω
, rad/s
|Hox|, -
10-1 100101
-360
-180
0
180 DI, CE output
ω
, rad/s
6Hox, deg
Fig. 12. Bode plots of the target and CE output dynamics: non-parametric estimates, model fits, and perfect target-tracking dynamics; single-subject data.
0
1
2
3
4
5
pursuit
preview
GN SI DI
Ke⋆, -
(a)
1
2
3
4
GN
Tl,e⋆, s
(b)
1
2
3
4
DI
TL,e⋆, s
(c)
0.15
0.20
0.25
0.30
0.35
GN SI DI
τ
v, s
(d)
4
8
12
16
10
10
SI DI
ω
nms, rad/s
(e)
0.00
0.25
0.50
0.75
10
10
SI DI
ζ
nms, -
(f)
Fig. 13. Estimated internal-error response (a-c) and physical limitation (d-f)
parameters. For GN and SI the NMS could be estimated for 0 and 10 subjects,
respectively.
offline model simulations, establishing the model’s capability
to predict HC behavior.
At the highest input frequencies, HCs cannot invert the CE
dynamics, as required to attain perfect target-tracking, with
just their far-viewpoint response. The role of the additive
near-viewpoint response is to better match the perfect target-
tracking dynamics at these high frequencies, and to further
increase the target crossover frequency. The hypothesized
0.0
0.5
1.0
1.5
measured
predicted
GN1 SI1 DI1
τ
f, s
(a)
0.0
0.5
1.0
1.5
GN1 SI1 DI1
Tl,f, s
(b)
0.0
0.5
1.0
1.5
GN1 SI1 DI1
Kf, -
GN0 SI0 DI0
(c)
Fig. 14. Estimated far-viewpoint parameters: negative delay
τ
f(a), lag time-
constant Tl,f(b), and gain Kf(c). Gray bars represent the individual subjects.
low but non-zero values for Kn(H.IV) were found in the
experiment for most subjects, but these were not always
supported by a clearly visible near-viewpoint response in the
corresponding non-parametric target response Hot(j
ω
). The
estimated value of Knis a poor indicator for the presence
of a near-viewpoint response, hence we cannot confidently
confirm H.IV. The near-viewpoint response varies substantially
between subjects, likely because it can yield only a marginal
performance benefit, at the cost of substantially more control

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 10
0.0
0.1
0.2
0.3
meas.
pred.
GN1 SI1 DI1
Kn, -
(a)
0.0
0.3
0.6
0.9
1.2
GN1 SI1 DI1
τ
n, s
(b)
Fig. 15. Estimated near-viewpoint parameters: gain Kn(a), and negative delay
τ
n(b). Gray bars represent the individual subjects.
activity. For adequate task performance, the far-viewpoint
response is much more important than the near-viewpoint
response.
To better illustrate human adaption between pursuit and
preview tracking tasks, and to the CE dynamics, we now
propose a first set of verbal adjustment rules. 1) Similar as in
compensatory tracking tasks [2], HCs equalize their internal
error response Hoe⋆(j
ω
)to the given CE dynamics such
that their combination exhibits integrator-like dynamics. 2) In
pursuit and preview tasks, HCs apply feedforward control by
adapting the relative target-tracking/CE-stabilization priority
through Kf, with more emphasis on target tracking (higher
Kf) in tasks with lower-order CEs. 3) In preview tasks,
HCs anticipate the target signal’s changes by basing their
“pursuit” response on the far viewpoint
τ
fs ahead, which
is positioned farther ahead for higher-order CEs. Hereby,
the response phase (hence performance) improves at lower
frequencies but deteriorates at higher frequencies. 4) HCs filter
these high frequencies from the previewed target signal by
adapting Tl,f; they filter away more high frequencies (higher
Tl,f) for higher-order CEs. 5) Optionally, performance can be
enhanced slightly more by also tracking the target signal’s high
frequencies with an additive, parallel near-viewpoint response,
which ideally resembles the inverse of the CE dynamics. With
a near-viewpoint response, HCs sacrifice some phase margin
in favor of a higher crossover frequency. These proposed
adjustment rules can be refined and extended by quantifying
HC adaptation to other task variables, like the preview time
and the forcing functions’ characteristics.
VII. CON CL US IO N
In this paper, we explained how humans use preview for
control in tracking tasks with various controlled element
dynamics. We presented offline analyses with a quasi-linear
model and results from a human-in-the-loop experiment, to
established the roles of the human’s near- and far-viewpoint
responses. Preview allows humans cancel their own and the
controlled element’s lags, up to relatively high frequencies, by
basing their far-viewpoint, pursuit response on the target signal
ahead; this requires no additional control activity. The optional
open-loop near-viewpoint response helps to synchronize the
output with the target signal at higher frequencies, but at
the cost of substantially more control activity. Target-tracking
performance improves primarily due to the far-viewpoint re-
sponse mechanism, while the benefit from the near-viewpoint
response is small. The adopted control-theoretic approach
provided unique quantitative insights into human control adap-
tation in preview tasks, which can explain human behavior
observed in other preview control tasks, like driving.
REF ER EN CE S
[1] L. R. Young, “On Adaptive Manual Control,” IEEE Trans. Man-Machine
Systems, vol. 10, no. 4, pp. 292–331, Dec. 1969.
[2] D. T. McRuer and H. R. Jex, “A Review of Quasi-Linear Pilot Models,”
IEEE Trans. Human Factors in Electronics, vol. 8, no. 3, pp. 231–249,
May 1967.
[3] D. T. McRuer, D. Graham, E. S. Krendel, and W. J. Reisener, “Human
Pilot Dynamics in Compensatory Systems, Theory Models and Exper-
iments with Controlled Element and Forcing Function Variations,” Air
Force Flight Dynamics Laboratory, Wright-Patterson Air Force Base,
OH, Tech. Rep. AFFDL-TR-65-15, 1965.
[4] W. W. Wierwille, J. R. Knight, and G. A. Gagne, “An Experimental
Study of Human Operator Models and Closed-Loop Analysis Methods
for High-Speed Automobile Driving,” IEEE Trans. Human Factors in
Electronics, vol. 8, no. 3, pp. 187–201, Sep. 1967.
[5] C. C. MacAdam, “Application of an Optimal Preview Control for
Simulation of Closed-Loop Automobile Driving,” IEEE Trans. Systems,
Man, and Cybernetics, vol. 11, no. 6, pp. 393–399, Jun. 1981.
[6] S. D. Keen and D. J. Cole, “Bias-Free Identification of a Linear
Model-Predictive Steering Controller From Measured Driver Steering
Behavior,” IEEE Trans. Systems, Man, and Cybernetics - Part B:
Cybernetics, vol. 42, no. 2, pp. 434–443, Apr. 2012.
[7] A. E. Patla and S. Rietdyk, “Visual Control of Limb Trajectory over
Obstacles during Locomotion: Effect of Obstacle Height and Width,”
Gait & Posture, vol. 1, pp. 45–60, 1993.
[8] R. A. Scheidt and C. Ghez, “Separate Adaptive Mechanisms for Con-
trolling Trajectory and Final Position in Reaching,” Journal Neurophys-
iology, vol. 98, pp. 3600–3613, 2007.
[9] L. D. Reid and N. H. Drewell, “A Pilot Model for Tracking with
Preview,” in Proc. 8th Ann. Conf. Manual Control, Ann Arbor, MI,
1972, pp. 191–204.
[10] M. Tomizuka and D. E. Whitney, “The Preview Control Problem with
Application to Man-Machine System Analysis,” in Proc. 9th Ann. Conf.
Manual Control, Cambridge, MA, 1973, pp. 429–441.
[11] K. Ito and M. Ito, “Tracking Behavior of Human Operators in Preview
Control Systems,” Electrical Eng. in Japan, vol. 95, no. 1, pp. 120–127,
1975, (Transl,: D.K. Ronbunshi, Vol. 95C, No. 2, Feb. 1975, pp 30-36).
[12] T. B. Sheridan, “Three Models of Preview Control,” IEEE Trans. Human
Factors in Electronics, vol. 7, no. 2, pp. 91–102, Jun. 1966.
[13] R. A. Hess, “A Model-Based Theory for Analyzing Human Control
Behavior,” in Advances in Man-Machine Research, W. B. Rouse, Ed.
JAI Press, 1985, vol. 2, pp. 129–176.
[14] K. van der El, D. M. Pool, H. J. Damveld, M. M. van Paassen,
and M. Mulder, “An Empirical Human Controller Model for Preview
Tracking Tasks,” IEEE Trans. on Cybernetics, vol. 46, no. 11, pp. 2609–
2621, Nov. 2016.
[15] R. J. Wasicko, D. T. McRuer, and R. E. Magdaleno, “Human Pilot
Dynamic Response in Single-loop Systems with Compensatory and
Pursuit Displays,” Air Force Flight Dynamics Laboratory, Wright-
Patterson Air Force Base, OH, Tech. Rep. AFFDL-TR-66-137, 1966.
[16] R. A. Hess, “Pursuit Tracking and Higher Levels of Skill Development in
the Human Pilot,” IEEE Trans. Systems, Man, and Cybernetics, vol. 11,
no. 4, pp. 262–273, Apr. 1981.
[17] M. C. Vos, D. M. Pool, H. J. Damveld, M. M. van Paassen, and
M. Mulder, “Identification of Multimodal Control Behavior in Pursuit
Tracking Tasks,” in Proc. 2014 IEEE Int. Conf. Systems, Man, and
Cybernetics, San Diego, CA, 2014, pp. 69–74.
[18] E. S. Krendel and D. T. McRuer, “A Servomechanics Approach to Skill
Development,” Journal of the Franklin Institute, vol. 269, no. 1, pp.
24–42, Jan. 1960.
[19] E. Donges, “A Two-Level Model of Driver Steering Behavior,” Human
Factors, vol. 20, no. 6, pp. 691–707, Dec. 1978.
[20] L. Saleh, P. Chevrel, F. Claveau, J. F. Lafay, and F. Mars, “Shared
Steering Control Between a Driver and an Automation: Stability in the
Presence of Driver Behavior Uncertainty,” IEEE Trans. on Intelligent
Transportation Systems, vol. 14, no. 2, pp. 974–983, 2013.
[21] H. R. Jex, R. E. Magdaleno, and A. M. Junker, “Roll Tracking Effects
of G-vector Tilt and Various Types of Motion Washout,” in Proc. 14th
Ann. Conf. on Manual Control, 1978, pp. 463–502.

IEEE TRANSACTIONS ON CYBERNETICS, VOL. XX, NO. XX, MONTH, YEAR 11
[22] M. M. van Paassen and M. Mulder, “Identification of Human Operator
Control Behaviour in Multiple-Loop Tracking Tasks,” in Proc. 7th
IFAC/IFIP/IFORS/IEA Symposium on Analysis, Design and Evaluation
of Man-Machine Systems, Kyoto, Japan, 1998, pp. 515–520.
[23] R. L. Stapleford, D. T. McRuer, and R. E. Magdaleno, “Pilot Describing
Function Measurements in a Multiloop Task,” IEEE Trans. Human
Factors in Electronics, vol. 8, no. 2, pp. 113–125, Jun. 1967.
[24] R. W. Allen and H. R. Jex, “An Experimental Investigation of Com-
pensatory and Pursuit Tracking Displays with Rate and Acceleration
Control Dynamics and a Disturbance Input,” NASA, Washington, D.C.,
NASA Contractor Report NASA CR-1082, 1968.
[25] G. C. Beerens, H. J. Damveld, M. Mulder, M. M. van Paassen, and J. C.
van der Vaart, “Investigation into Crossover Regression in Compensatory
Manual Tracking Tasks,” Journal of Guidance, Control, and Dynamics,
vol. 32, no. 5, pp. 1429–1445, Sep.-Oct. 2009.
[26] V. A. Laurense, D. M. Pool, H. J. Damveld, M. M. van Paassen, and
M. Mulder, “Effects of Controlled Element Dynamics on Human Feed-
forward Behavior in Ramp-Tracking Tasks,” IEEE Trans. Cybernetics,
vol. 45, no. 2, pp. 253–265, Feb. 2015.
Kasper van der El (S’15) received the M.Sc. degree
in aerospace engineering (cum laude) from TU Delft,
The Netherlands, in 2013, for his research on manual
control behavior in preview tracking tasks. He is
currently pursuing the Ph.D. degree with the section
Control and Simulation, Aerospace Engineering, TU
Delft. His Ph.D. research focuses on measuring and
modeling human manual control behavior in general
control tasks with preview. His current research in-
terests include cybernetics, mathematical modeling,
and system identification and parameter estimation.
Daan M. Pool (M’09) received the M.Sc. and Ph.D.
degrees (cum laude) from TU Delft, The Nether-
lands, in 2007 and 2012, respectively. His Ph.D.
research focused on the development of an objective
method for optimization of flight simulator motion
cueing fidelity based on measurements of pilot con-
trol behavior. He is currently an Assistant Professor
with the section Control and Simulation, Aerospace
Engineering, TU Delft. His research interests include
cybernetics, manual vehicle control, simulator-based
training, and mathematical modeling, identification,
and optimization techniques.
Marinus (Ren´
e) M. van Paassen (M’08, SM’15)
received the M.Sc. and Ph.D. degrees from TU Delft,
The Netherlands, in 1988 and 1994, respectively, for
his studies on the role of the neuromuscular system
of the pilot’s arm in manual control.
He is currently an Associate Professor at the
section Control and Simulation, Aerospace Engi-
neering, TU Delft, working on human-machine inter-
action and aircraft simulation. His work on human-
machine interaction ranges from studies of per-
ceptual processes and manual control to complex
cognitive systems. In the latter field, he applies cognitive systems engineering
analysis (abstraction hierarchy and multilevel flow modeling) and ecological
interface design to the work domain of vehicle control.
Dr. van Paassen is an Associate Editor of the I EEE TRAN SAC TIO NS ON
HUM AN- MAC HIN E SYS TEM S.
Max Mulder (M’14) received the M.Sc. degree and
Ph.D. degree (cum laude) in aerospace engineer-
ing from TU Delft, The Netherlands, in 1992 and
1999, respectively, for his work on the cybernetics
of tunnel-in-the-sky displays. He is currently Full
Professor and Head of the section Control and
Simulation, Aerospace Engineering, TU Delft. His
research interests include cybernetics and its use in
modeling human perception and performance, and
cognitive systems engineering and its application in
the design of “ecological” interfaces.