=Paper=
{{Paper
|id=Vol-1419/paper0082
|storemode=property
|title=Cognitive Control of Pacing During Endurance Exercise: Everyone is a Quitter
|pdfUrl=https://ceur-ws.org/Vol-1419/paper0082.pdf
|volume=Vol-1419
|dblpUrl=https://dblp.org/rec/conf/eapcogsci/BestW15
}}
==Cognitive Control of Pacing During Endurance Exercise: Everyone is a Quitter==
https://ceur-ws.org/Vol-1419/paper0082.pdf
Cognitive Control of Pacing During Endurance Exercise: Everyone is a Quitter
Bradley J. Best (bjbest@adcogsys.com)
Kristin M. Weinzierl (kristin.weinzierl@gmail.com)
Adaptive Cognitive Systems LLC, 229 Marine Dr.
Blaine, WA 98230 USA
Abstract can push themselves to the limits of physiology. A maximal
physical effort, if it is purely limited by bodily constraints,
Contemporary accounts of control of pacing during endurance
exercise focus on physical limitations, generally assuming has little need to involve cognition. A sufficiently motivated
humans work to that physical limit. Conceptually, control is participant is assumed to produce a maximal effort, and
ceded to the body at the beginning of an exercise bout and is cognition is reduced to a simple on/off switch, where it
returned to central cognition upon achieving a state of invokes the required effort.
exhaustion. We advance an alternative decision-making Pacing strategy is a matter of identifying the maximal
model of control of pacing, where the decision whether to power output that can be sustained across the expected time
persist in the effort is revisited continuously, and cessation of
the exercise bout is an explicit, cognitively controlled interval. For the elite athletes commonly used in pacing
decision. Our model depends on the following assumptions studies, the self selection of a maximal pace is done with
and features: 1) decisions are made in discrete cycles, 2) relative ease, resulting in models that have slight variations
repetitive bodily motions depend on a central pattern from constant power output, but which can be explained
generator, 3) afferent physiological feedback produces a sense through simple physiological explanations such as reserves
of perceived exertion, 4) central cognition mediates between of anaerobic energy. One commonly used model, known as
perceived exertion and the value of persisting (motivation) to
the Hill model (after A.V. Hill), is based on the idea that
perform an ongoing cost-benefit analysis, and 5) cessation of
exercise occurs when an explicit decision is made to exercise produces linear changes in metabolism, until
discontinue the effort. demand exceeds capacity, resulting in fatigue and cessation
of exercise. This simple, conventional model of exercise
Keywords: cognitive control of pacing; central pattern
generator; perceived exertion; cognitive models of exercise
performance also produces some surprising predictions,
however, which have been justifiably criticized (Noakes,
Cognitive Control of Pacing 2011). Among these suspect predictions are the existence of
a single maximal workload (regardless of distance or time),
Why do we continue in the face of fatigue, and when and and the inability to lift the pace at the end of a “maximal”
why do we give up? This question is so fundamental, an session. The ubiquity of a mad sprint for the line in long
entire body of research in the field of exercise physiology distance endurance events falsifies the prediction outright.
has been dedicated to answering it. Much of the field has That the Hill model has survived nearly a century of
focused on physical determinants of eventual performance, application is a testament to its utility in explaining some
whether it be maximal oxygen consumption, hydration, or important phenomena, and to the lengths to which
mechanical factors such as leg length or body composition. experimenters have succeeded in removing the brain of the
Motivational factors are often assumed away, frequently by experimental participant from the experiment.
studying world-class athletes who can be broadly
characterized as exceptionally motivated, and cognitive Putting the Brain Back in Charge of Pacing
factors are easily neglected. The physical performance is
treated much as one would test an internal combustion The trend toward removing cognitive aspects of “exercise to
engine, with maximal performance determined by the fatigue” has been turned on its head in several studies,
physical properties of the engine. That these athletes would however, where deception about pacing has been used to
persist during testing is a given, and that they produce a examine the cognitive inputs to sustained physical exertion.
maximal effort is one of the further assumptions that defines For example, (Stone et al., 2011) conducted a study in
the implications of the studies. Cognition is given short which participants completed a cycling time trial (a timed
shrift, both in terms of theory and methodology. solo effort across a fixed distance) in a simulation
Common protocols consist of exercising to exhaustion at environment against an avatar that represented their own
a specified intensity, or completing a set distance in a best prior performance (supposedly a maximal effort). The
minimal time. Pacing models have been proposed to critical manipulation was a deception condition, where that
describe human behavior; these are predicated on an often prior performance was augmented with a 2% increase in
hidden assumption that physiology determines pacing (via power output. Participants, believing that performance
physical fatigue), rather than that behavior drives pacing represented something they had already done, consistently
(via cognition). This assumption is rarely exposed in outperformed the deceptive performance. The researchers
constant load exercise because experimenters have worked concluded that participants all had a metabolic reserve that
to remove cognition from the performance and isolate the they strategically conserved, and thus none had completed a
physical aspect, often through the use of expert athletes who truly maximal effort during their initial best efforts. From
505
this perspective, pacing strategy reflects cognitive budgeting appearance of near-optimality without reference to a
of available resources against anticipated demands. preconceived plan.
Tucker (2013) argues that pacing is the application of a In the remainder of this paper we will 1) formalize this
plan, the entirety of which the participant is not completely theory within the framework of a cognitive architecture, and
aware of, to spend available resources to achieve the goal in 2) demonstrate its utility through a computational model of
a near-optimal fashion (where the difference between exercise pacing.
optimal and failure is often less than 1%). He defines pacing
during exercise as an attempt to optimally meet the Using a Unified Theory of Cognition to Constrain a
following goals and constraints: Theory of Cognitive Control of Pacing
1. use available energy at the optimal rate Unified theories of cognition (UTCs; Newell, 1994) attempt
2. gain heat slowly enough to complete the task, but to collect the invariants of human cognitive behavior within
not so slowly as to reduce intensity a single, computationally realizable framework. One of the
3. accumulate metabolites at a low enough rate to primary benefits of depending on UTCs is the requirement
avoid being overwhelmed by them to make process models explicit and comprehensive. We
4. meet oxygen requirements of muscle, brain and turn to one candidate UTC, the ACT-R cognitive
other tissues architecture (Anderson et. al., 2004), as a source of
5. compete with other runners, the clock or whatever structural constraints on cognitive processing to inform the
other motivational factors impact on performance development of a theory of cognitive pacing. Critical to this
The conceptual model of Tucker (2013) depends on a paper, ACT-R has also been mapped onto a variety of brain
template matching process that occurs continuously, areas, and can be used to predict and explain brain activity
weighing task demands against these templates for during task performance. A core tenet of ACT-R is that
performance. Tucker (2013) additionally posits that pacing central cognition can be very finely approximated using a
differences are due to “uncertainty” about the interpretation discrete decision-making cycle., with a pattern matching
of templates in the context of the task, which results in the system implemented as a production system (that maps onto
maintenance of a metabolic reserve. What this model lacks, the basal ganglia) performing a repetitive decision-making
however, is any specificity or concrete definition of what inner loop during task performance. The central decision-
these templates might be, or what the uncertainty is and how making process interacts on each cycle with peripheral
it is applied. systems such as memory, visual, auditory, and haptic,
It is exactly those theoretical gaps that we intend to perception, and bodily motor functions through a set of low
address here, by making our model computationally explicit. capacity interfaces, or buffers, which allow limited
We assert that, while participants may have a rough goal to parallelism.
do their best, they are engaged in an ongoing comparison When running or cycling, ~150-200 individual leg
between the expected duration of the work bout and their movements are typically made per minute. This
current intensity of effort with reference to their prior automaticity requires a helper system capable of regulating
experiences. That is, they are retrieving prior events from repetitive motions without the need to burden central
memory for comparison, where the content of this memory cognition. That is, without such a helper system, one would
includes aspects such as effort, duration, environmental be unable to do anything other than perform the exercise
conditions, and, critically, sustainability of the effort. itself because there would be no free cycles to devote
The participant need not balance, nor even be aware of, anywhere else. Thus, it is apparent, even without recourse to
most of the factors that define endurance performance a UTC-based analysis, that because endurance exercise does
during the majority of work bouts. The surprising not overwhelm central cognition, it follows that it must
concordance of physiological limitations (where body primarily be handled elsewhere.
temperature, energy reserves, and cardiac output, for Turning back to UTCs, modeling highly interactive real-
example, simultaneously reach their limits during work to time tasks often requires helper systems running at higher
exhaustion) can be explained largely by physiological frequencies than central cognition. For example, Best and
adaptations: systems that fail often adapt first, until all are Lebiere (2003) were only able to demonstrate smooth
roughly on par with each other. For example, the ability to targeting, object tracking, and movement behavior in a
handle heat stress can change dramatically with only a few virtual environment by reducing the cycle time to ~10ms,
weeks of training. There is no need for the athlete to attempt violating the fundamental of the ACT-R cognitive
to optimize these factors individually, much less be aware of architecture (“overclocking” central cognition). Salvucci et
them in many cases. On the other hand, a participant is al. (2001) addressed this limitation in a driving task by
likely to be acutely aware of any single system (whether modifying the core architecture to interact with a slave
body cooling, oxygen delivery, or bodily afferent feedback system running at a higher frequency, thereby respecting the
such as muscular pain) that signals an impending or realized constraints ACT-R places on central cognition.
failure. Thus, by reacting to the system that corresponds to The implication of this analysis is that, in the context of
the weakest link and matching the current effort level based endurance exercise, there must be a “helper” that conducts
on personal history, cognitive control can give the and regulates much of the activity involved in endurance
506
exercise, and central cognition can be expected to primarily • Central Executive needs to increase effort signal to
interact with that helper. CPG as fatigue occurs to maintain the same muscular
One candidate slave system that might provide a link input
between central cognition and endurance exercise is the • Pacing should not be natural, but should emerge with
Central Pattern Generator, a spinal network capable of learned experience over time
producing rhythmic limb movements in the absence of • With experience, cadence will tend toward just under
cognitive control (Dimitrijevic et al., 1998). The CPG tends 90, but while learning it will be lower. This is not
toward a natural resonant frequency of just under 3Hz, automaticity or power law speedup, but rather a
which corresponds to a natural cadence of ~90, matching up removal of the cognitive effort of deliberative
closely with observed freely selected cadences of runners processing that is replaced with the natural frequency
and cyclists. This proposed model of control thus provides a (speedup beyond a cadence of 90 should not happen
level of indirection between central cognition and the with greater experience).
exercising muscles. Fatigue signals from the muscles We will next examine whether these predictions are
operate directly on CPG, which then passes this information sustained or contradicted by the existing literature, and we
on to the central executive. An effort signal from the Central will explore a computational model that implements this
Executive causes firing, but fatigue of neural pathways will model, providing a proof of the theoretical concept.
cause reduced output for the same input firing signal.
Within the central executive, this model of control is Constraints of Human Physiology on Endurance
based on the retrieval of relevant templates and comparison The preceding discussion focuses on the cognitive control
to ratings of perceived exertion (RPE). Given a target of pacing. There are also hard limits on endurance
time/distance, memory can be scanned for a relevant effort performance imposed directly by human physiology. An
that was successfully made at that time/distance. Given individual's capacity to perform endurance exercise is
ongoing RPE feedback, the effort can be increased or characterized by many features; chief among these are:
reduced relative to the current effort. • Aerobic capability: the ability to metabolize oxygen to
Using this model, undershooting and overshooting of
produce work, at lower intensities and long time scales.
pacing efforts are both possible and likely. Overshooting has
This ability may be defined in terms of critical power
worse outcomes (failure), while undershooting can result in
(the maximal work rate that can be indefinitely
less than optimal performance. Specifically, undershooting
sustained), and is commonly expressed in units of
leaves energy to be spent more rapidly at the end (end
oxygen consumed per measure of body weight.
spurt), but due to task limitations, it may not be possible to
• Anaerobic work capacity (AWC): the conversion of
spend all of the available energy. In all cases, these
stored chemical energy to work without the use of
experiences are learned and stored, resulting in accumulated
oxygen), at higher intensities and shorter time scales.
knowledge with experience. In the absence of experience,
This anaerobic work creates an oxygen debt that must
pacing can be expected to fail often, since there are no
be repaid through respiration.
successes to draw from. This naturally produces learning
• Heat tolerance: the ability to maintain homeostasis in
predictions as well, since lack of experience should result in
many more failures in pacing. response to heat produced through exercise, primarily
through sweat production and diversion of blood to
Specific Theoretical Predictions The constraints discussed surface skin capillaries to radiate heat.
above result in the following theoretical predictions: • Maximal power output: the greatest work rate that can
• Completely inexperienced athletes (as young athletes be sustained regardless of the timespan.
often are) are likely to have more failures of over- • Muscular fatigue: the generation of chemical waste
pacing and under-pacing products that inhibit further muscular action.
• Experienced athletes, after learning to avoid failures, While there are many other factors that influence the
will pace conservatively, and will need to spend more ability to perform endurance exercise, these factors capture
of their energy budget at the end of the effort, many of the main constraints that impact pacing. Aerobic
producing an end spurt (increase in output intensity capability and anaerobic work capacity are the two factors
near the end of a work bout) used in the critical power model of Monod and Scherrer
• Attentional manipulations that divide attention will (1965), which predicts the maximal duration Tlim of
endurance exercise as a function of work rate P, the
negatively impact pacing (often through a reduction in
anaerobic work capacity AWC, and the work rate that can be
cadence that reduces work rate)
sustained indefinitely CP. Their relationship is given by:
• Attentional manipulations that focus attention will
positively impact pacing (through maintenance of an Tlim=AWC/(P-CP)
even pacing strategy)
• Perceived Effort interacts by way of interruption,
focuses attention on perception of effort (pain,
discomfort, effort)
507
Graphically, the CP model traces an asymptotic current context, producing an assessment of whether a
hyperbolic curve, predicting work rates that approach particular effort will succeed or fail, weighted toward recent
infinity as the time approaches 0, and work rates that never memories (that is, exhibiting recency). Thus, we might
decrease beyond asymptote as time approaches infinity. expect athletes after a layoff period of reduced or no
Despite this limitation, the CP model provides an excellent training to overestimate the appropriate pace, because their
account of performance from durations of several minutes to last memory corresponded to a higher level of performance.
several hours. Figure 1 below depicts the CP model in In our model, this matching process is a noisy, inexact
relation to actual historical performances for one individual match to prior memory, which might be influenced by
athlete, showing the ability of this model to predict context, recency, and other similar factors known to
limitations while taking individual differences into account. influence memorability.
The conventional orient-decide-act cycle, embedded
1200 within cognitive architectures, is also a critical component
1000
of our model. As we previously pointed out, if endurance
exercise required constant attention, it would overwhelm
Power Output (W)
800 attentional resources. The central pattern generator may
CP Model largely be responsible for handling the continuance of
600 Performance
exercise under non-challenging circumstances. The critical
400 question relates to when and how often attention focuses the
individual on reconsidering the decision to either persist,
200
0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
increase, decrease, or entirely cease the effort, during
Time: Log(seconds)
challenging efforts.
Fatigue and pain are both implicated in this refocusing of
Figure 1: Historical time-intensity endurance curve for an attention. Anyone who has ever engaged in repetitive
individual athlete. The solid line shows the prediction of the physical activity knows that the activity will become more
CP model of Monod and Scherrer (1965) across the range of difficult during a work bout, no matter the intensity. One
its practical applicability; the triangular markers show the important reason for this is the fatigue of neuromuscular
athlete's actual historical performances. circuits. Specifically, achieving the same muscular output
(measured via EMG) requires stronger and stronger neural
Our modeling goal is to predict these work rate – duration input, which corresponds to an increasing perception of
curves through a model of cognitive control, given a set of effort. Voluntary muscle activation, which is measured as
inputs available to the cognitive system. This requires the percentage of neural activity achieved under voluntary
interaction with a physiology module that incorporates the control when compared to direct electrical stimulation, is
modeling of heat generation and dissipation, oxygen approximately 90% prior to fatigue, but drops to less than
consumption (as a function of resting metabolism, aerobic 75% under conditions of fatigue. This mechanism protects
exercise, and repayment of oxygen debts), anaerobic energy muscle from catastrophic damage, regulating exercise by
use and reserves, and the provision of a perceived exertion adjusting output from the brain (Amann, et al, 2008).
signal. This last quantity, perceived exertion, is primarily Exercising muscles also experience microscopic damage
based on cardiac output (the original RPE scale, in fact, was and produce metabolic waste products, leading to a
a linear transformation of heart rate), but also includes heat perception of muscular pain and fatigue. This ever-
stress, and muscular fatigue components (Borg, 1982). We increasing pain signal further attracts attention. The neural
have implemented such a module, allowing us to derive fatigue protective mechanism also interacts with regulation
oxygen consumption, heart rate, heat generation, and of activity via the CPG indirectly. At exactly the time when
anaerobic energy status from a particular work rate given an an athlete needs to focus attention on making the decision to
athlete's individual physiological parameters. We now turn maintain an effort (to increase the neural output signal to the
to the model of control. exercising muscles), they may be interrupted to process the
urgent sensory input of a pain signal.
Cognitive Control of Pacing Behavior Finally, while the CPG might allow the repetition of
The preceding discussion outlines the effort-duration rhythmic activities, it will not drive effortful performance on
relationship for endurance exercise. The cornerstone of our its own. In the absence of a decision to continue, CPG-
model is the use of memory for historical efforts as the basis driven behavior will revert to the natural resonant frequency
for establishing and refining a current effort. Those of the CPG, possibly reducing cadence during high effort
memories, or their absence, are exactly the source of times, and the same CPG signal will result in a decreasing
predictions of expert-novice differences. Given a specific output effort due to fatigue of neural pathways, despite any
duration, an athlete will gauge their effort based on conscious decision to continue.
historical experience. Specifically, we suggest that a The decision to persist during endurance exercise to
blended memory retrieval (Anderson et al., 2004) is exhaustion, thus, must be revisited more and more
performed, which combines prior experiences similar to the frequently as the interval continues through the involuntary
508
509
2011). Indeed, these theories treat the exercising individual Pegelow D.F., Dempsey J.A. (2008). Somatosensory
much like a solid rocket – once ignition is achieved, the feedback from the limbs exerts inhibitory influences on
rocket burns until the fuel is gone, often at the ideal rate for central neural drive during whole body endurance
the event duration. While this may seem silly, the ability of exercise. J Appl Physiol 105: 1714–1724.
athletes to finely gauge their performances to match the Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S.,
limit of their endurance often masks the role of cognition, Lebiere, C., & Qin, Y . (2004). An integrated theory of
allowing this description to capture the rough shape of the mind. Psychological Review 111, (4). 1036-1060.
behavior. It is not until we look at inexperienced athletes Best, B. J., & Lebiere, C. (2003). Teamwork,
that we must account for the import of learning, and thus communication, and planning in ACT-R agents engaging
cognition, in pacing behavior. in urban combat in virtual environments. In Proceedings
Our model, though simple, accomplishes the following: of the 2003 IJCAI Workshop on Cognitive Modeling of
• Defines a conceptual theoretical model of cognitive Agents and Multi-Agent Interactions (pp. 64-72).
control of pacing Borg, G. A. (1982). Psychophysical bases of perceived
• Defines the interaction of central cognition with a exertion. Medicine & Science in Sports &
central pattern generator as a mechanism for rhythmic Exercise;14(5):377-81.
exercise. Dimitrijevic, M.R., Gerasimenko, Y., Pinter, M.M. (1998).
• Defines the role of memory of prior efforts in Evidence for a spinal central pattern generator in humans.
determining appropriate efforts. Ann. N. Y. Acad. Sci. 860 (1 Neuronal Mech): 360–76.
• Predicts developmental failures in pacing and, in doi:10.1111/j.1749-6632.1998.tb09062.x. PMID 9928325
particular, oscillations between excessive and Gonzalez, C., Best, B. J., Healy, A. F., Kole, J. A., &
insufficient effort to maximize performance. Bourne, L. E. (2011). A cognitive modeling account of
• Predicts improvements in pacing performance with simultaneous learning and fatigue effects. Cognitive
accumulation of memories of prior successes and Systems Research, 12, pp. 19-32. Elsevier.
failures of pacing to draw from. Monod, H., & Scherrer, J. (1965). The work capacity of a
• Predicts a conservative undershoot heuristic to avoid synergic muscle group. Ergonomics, Vol 8(3), 1965, 329-
failure of the effort. 338. http://dx.doi.org/10.1080/00140136508930810
• Predicts an end-spurt phenomenon as a means of Newell, A. (1994). Unified Theories of Cognition, Harvard
University Press; Reprint edition.
spending an energy budget after undershooting
Noakes, T. D. (2011). Time to move beyond a brainless
To our knowledge, this is the first model of endurance
exercise physiology: the evidence for complex regulation
exercise pacing to situate pacing as a series of ongoing
of human exercise performance. Applied Physiology
decisions within a broader cognitive framework capable of
Nutrition Metabolism; 36(1):23-35. doi: 10.1139/H10-
performing other cognitive tasks. As such, it represents a
082.
bridge between two disparate research communities. This
Salvucci, D. D., Boer, E. R., & Liu, A. (2001). Toward an
disparity presents the corresponding challenge of sharing
integrated model of driver behavior in a cognitive
ideas in the absence of a shared common vocabulary to
architecture. Transportation Research Record,1779, 9-16.
describe the phenomenon. This paper is one step toward
Stevens, S.S. (1957). "On the psychophysical law".
establishing some of that shared context within the cognitive
Psychological Review; 6 4 ( 3 ) : 1 5 3 – 1 8 1 .
science community, connecting cognitive behavior back to
doi:10.1037/h0046162
the physical world of human physiology.
Stone, Mark, Thomas, Kevin, Wilkinson, Mick, Jones,
Our future efforts will also explore modeling incomplete
Andrew, St Clair Gibson, Alan and Thompson, Kevin
knowledge and attempt to capture the developmental trends
(2011). Effects of deception on exercise performance:
more clearly. When presented with few or no relevant
implications for determinants of fatigue in humans.
instances of prior behavior, the current model easily ends up
Medicine & Science in Sports & Exercise, 44 (3). pp. 534-
in a failure state, where it is unable to complete the effort, or
541.
unable to settle on a steady state that corresponds to a
Tucker, R. (2009). The anticipatory regulation of
sustainable effort. While we have not yet attempted to
performance: the physiological basis for pacing strategies
model that learning phenomenon, our initial explorations
and the development of a perception-based model for
suggest that this platform will provide an excellent path
exercise performance. British journal of sports medicine.
forwards to modeling the learning of pacing behavior.
http://www.ncbi.nlm.nih.gov/pubmed/19224911
Finally, what the model presented here lacks, in its
Tucker, R. (2013). Fatigue, invisible Barriers, physiological
current form, are a complete instantiation within a cognitive
limits and performance: the role of the brain in
architecture, and a thorough validation against rich data sets.
performance psychology. Talk presented at the Marathon
Our future efforts will be concentrated in this direction.
Medicine 2013 Conference, London, England.
Zwislocki, J. J. (2009). Sensory Neuroscience: Four Laws of
References Psychophysics: Four Laws of Psychophysics. Springer.
Amann M., Proctor L.T., Sebranek J.J., Eldridge M.W.,
510