MODELING FATIGUE IN A COGNITIVE ARCHITECTURE: WHAT IS THE VALUE ADDED 26 October 06

1
MODELING FATIGUE IN A COGNITIVEARCHITECTURE
WHAT IS THE VALUE ADDED? 26 October 06

• Glenn Gunzelmann
• Kevin Gluck
• Human Effectiveness Directorate
• Air Force Research Laboratory

AFOSR Grant 04HE02COR
2
Acknowledgments

• Funding
• AFOSR Grant 04HE02COR
• Willard Larkin
• Air Force Research Laboratory (AFRL)
• Warfighter Readiness Research Division
• Data
• University of Pennsylvania, Division of Sleep and
  Chronobiology
• David Dinges/(Hans Van Dongen)/Robert OConnor
• Conversations
• Performance and Learning Models (PALM) Team
  (AFRL)
• Frank Ritter (PSU)
• Beth Klerman (Harvard/BWH)
• Hans Van Dongen/Greg Belenky (WSU)
• David Dinges (UPenn)

3
Outline

• Background and Approach
• Identifying mechanisms in the architecture
• What do cognitive models tell us?
• Conclusions

4
Applications of Fatigue Models
5
Objective

• A priori predictions of the effect of sleep loss
  and circadian rhythms on human cognition and
  performance in complex, novel tasks

Approach
Use a cognitive architecture to bridge the gap
between biomathematical predictions of alertness
and human cognition and performance
6
Biomathematical Models

• Biomathematical models capture the dynamics of
  alertness
• Interaction of
• Circadian rhythms
• Mechanisms of the suprachiasmatic nucleus (SCN)
• Sleep homeostat
• Mechanisms of the ventrolateral preoptic area
  (VLPO)
• Human performance is used as a window into this
  interaction
• SCN VLPO
• Not where computations for task performance are
  done
• Do not direct visual attention or plan/execute
  motor actions
• Do not calculate results for addition facts
• Influence information processing in other areas
• Cliff Sapers mention of projections to the
  thalamus cortex
• Moderate cognitive performance

7
Cognitive Architectures

• Research Perspective (Newell, 1990)
• A single system (mind) produces all aspects of
  behavior it is necessary to have a theory that
  provides the total picture and explains the role
  of the parts and why they exist.
• Circadian rhythm sleep homeostat are two
  parts
• Cognitive architectures contain mechanisms for
  other parts and how they interact
• Central cognition, control, memory, perception,
  action,
• Implemented in software
• Cognitive mechanisms operate to generate behavior
  and performance predictions

8
How the Modeling Efforts Fit Together

• Biomathematical models
• Mechanisms of circadian cycle sleep homeostat
• Computational cognitive models
• Mechanisms of human information processing
• Integration
• Identify the impact of fatigue on cognitive
  processes
• Why how does processing in SCN/VLPO impact
  cognitive performance?
• The cognitive impact of the links between the
  arousal system and the thalamus cortex
• The bottom-up view that David Dinges mentioned
• Predict changes in performance in novel tasks

9
Outline

• Background and Approach
• Identifying mechanisms in the architecture
• What do cognitive models tell us?
• Conclusions

10
We Are Using ACT-R
Atomic Components of Thought - Rational
Anderson et al., 2004

• A serial production system
• Central cognition
• Distinct modules represent different processing
  subsystems
• Perception, motor action, declarative memory,
• Serial processing within modules
• But activity is parallel across modules
• Correspondence to the brain
• Mechanisms and modules mapped to brain areas
• Predicts fMRI BOLD response
• Neurally-inspired subsymbolic mechanisms
• Moderate behavior of the symbolic level
• Generates observable behavior

11
Explanatory Breadth

• Perception and Attention
• Psychophysical Judgements, Visual Search, Eye
  Movements, Multi-Tasking, Task Switching,
  Subitizing, Stroop, Driving and Flying Behavior,
  Situational Awareness, Embedded Cognition,
  Graphical User Interfaces, Time Perception
• Learning and Memory
• List Memory, Interference, Implicit Learning,
  Skill Acquisition, Cognitive Arithmetic, Category
  Learning, Learning by Exploration and
  Demonstration, Updating Memory and Prospective
  Memory, Causal Learning, Working Memory, Practice
  and Retention, Representation
• Language Processing
• Parsing, Lexical Processing, Analogy and
  Metaphor, Sentence Memory, Language Learning
• Problem Solving and Decision Making
• Tower of Hanoi, Choice and Strategy Selection,
  Mathematical Problem Solving, Spatial Reasoning
  and Navigation, Dynamic Systems, Use and Design
  of Artifacts, Game Playing, Insight and
  Scientific Discovery, Programming, Reasoning,
  Errors
• Other
• Cognitive Development, Information Search,
  Cognitive Workload, Individual Differences,
  Motivation, Emotion, Cognitive Moderators,
  Cognitive Workload, Computer Generated Forces,
  Video Games and Agents, fMRI, Communication,
  Negotiation Group Decision Making, Instructional
  Materials, User Modeling, Intelligent Tutoring
  Systems

From http//actr.psy.cmu.edu
12
Mechanisms 1 Central Cognition

• Task Psychomotor Vigilance Task (PVT)
• Sustained attention task
• Wait for a stimulus to appear (delay varies from
  2 to 10 seconds)
• Respond by pressing a button when it does
• A session lasts for 10 minutes
• Sensitive to levels of sleep deprivation and
  circadian desynchrony
• Very close to the architecture
• Emphasizes basic information processing
  mechanisms
• In contrast to knowledge-intensive tasks
• No learning curve in human performance
• Median reaction time is around 250 ms for rested
  people

13
PVT 88 Hours Without Sleep
Empirical data from Van Dongen et al., 2001
Human Data

• Mechanism selected based on neurobehavioral data
  combined with theoretical mapping of ACT-R
  mechanisms to brain regions
• Implicates the thalamus and basal ganglia
• Based upon the model completing 10-minute PVT
  sessions
• Does the task, just like human participants all
  DVs simultaneously

14
Integrate Biomathematical Models

• Theoretical basis for predicting alertness
• These predictions drive parameter changes
• Include two different models
• Circadian Neurobehavioral Performance and
  Alertness
• Harvard Jewett Kronauer, 1999
• Sleep, Alertness, Fatigue, Task Effectiveness
• AFRL, SAIC Hursh et al., 2004
• Josh Gross, a graduate student at Penn State,
  contributed to this work

15
Model Performance
Empirical data from Van Dongen et al., 2001
Lapses
16
Mechanisms 2 Declarative Knowledge

• Serial Addition/Subtraction Task
• Two single-digit numbers operator (/-)
  presented
• Perform operation (1st ltoperatorgt 2nd)
• 3 6 ? 3 6
• Response
• If the result is positive ? ones digit
• If the result is negative ? Result 10
• Impact of increased fatigue
• Decreased accuracy
• Increased response times

17
ACT-R Based Account

• New task model created to perform SAST
• Includes declarative procedural knowledge
• Addition/subtraction facts (declarative)
• Encoding solution process (procedural)
• Procedural mechanisms developed for PVT are
  inadequate
• Illustrates that fatigue impacts multiple aspects
  of cognition
• Declarative memory has parallel mechanisms
• Weve added parameters from these mechanisms to
  our account
• Broadens our account of how fatigue impacts
  cognition

18
Model Performance 88 Hrs. Awake
Empirical data from Van Dongen et al., 2001

• Model is able to capture changes in both accuracy
  and response times
• Mechanisms are theoretically motivated
• Parallel mechanisms identified for procedural
  knowledge
• Produces a more comprehensive set of mechanisms
  impacted by fatigue in ACT-R
• Reflects the global impact of fatigue on
  cognitive performance

19
Outline

• Background and Approach
• Identifying mechanisms in the architecture
• What do cognitive models tell us?
• Conclusions

20
What Weve Shown So Far

• Ability to capture impact of fatigue on
  performance
• Mechanisms in the architecture
• Combined with knowledge for particular tasks
• Using biomathematical models to drive parameter
  changes
• Fit of cognitive models is comparable to
  biomathematical models
• Level of explanation is different
• Biomathematical models reflect physiological
  changes scaled to performance data
• Cognitive models reflect changes in information
  processing resulting from fatigue

21
Cognitive Models Can Tell Us Why

• Performance changes in PVT
• Increased false starts
• Decreased alertness decreases discriminability of
  alternatives that differ in likelihood of success
• Increased median RT
• Decreased alertness decreases the probability
  that any action will exceed the utility threshold
• Increased occurrence of micro-lapses
• ACT-R responds optimally less often
• Increase in lapses sleep attacks
• Progressive reduction in alertness increases
  likelihood further
• Longer sequences of micro-lapses
• ACT-R falls asleep

22

• Performance on SAST
• Increased response times
• Decreased activation of declarative information
• Increased retrieval times
• Decreased accuracy
• Increased retrieval times result in failures to
  encode problem elements
• Encoding one item is not completed until next
  item is already gone
• Retrieval failures increase likelihood of
  retrieval errors

23
Novel Predictions

• Cognitive models interact with the same software
  as human participants
• They do the task
• Possible to ask new questions
• SAST involves different types of problems
• Result of the operation is positive
• Result of the operation is negative
• More difficult Additional operations needed
• Model makes predictions about impact of fatigue
  on these different problems
• Empirical data is not available
• Biomathematical models alone cannot do this

24
Sample Predictions Response Time
Detailed data on human performance are critical!!
r(Diff,CNPA) -0.83 r(Diff,SAFTE) -0.79
25
What about Novel Tasks?

• Applications of models of fatigue involve
  complex, dynamic tasks
• Difficult or impossible to get empirical data
  about impact of fatigue
• Mechanisms in cognitive architectures apply
  across tasks and domains
• The same is true for the impact of fatigue on
  those mechanisms
• By developing models for novel tasks, a priori
  predictions can be made about the impact of
  fatigue
• Biomathematical models alone cannot do this

26
Conclusions

• Demonstrated ability to capture human performance
  in multiple tasks
• Utility of a cognitive architecture for
  identifying how fatigue impacts performance
• Combined explanatory power of biomathematical
  models of alertness linked to a cognitive
  architecture
• Detailed account of the dynamics of fatigue
• Validated theory of human information processing
  mechanisms
• Value of computational cognitive models for
  making predictions
• Extending predictions for existing tasks
• Potential for a priori predictions for novel tasks

27
Questions?
Thank You