Sie sind auf Seite 1von 5

PROCEEDINGS of the HUMAN FACTORS AND ERGONOMICS SOCIETY 46th ANNUAL MEETING -- 2002 303

Proceedings of the Human Factors and Ergonomics Society 46th Annual Meeting -- 2002

ONCE MORE WITH FEELING:


AUGMENTING RECOGNITION PRIMED DECISION
MAKING WITH AFFECTIVE FACTORS

Eva Hudlicka & Jonathan Pfautz*


Psychometrix Associates, Inc., Blacksburg, VA
*Research Laboratory of Electronics, MIT, Cambridge, MA

Although quintessentially human, emotions have, until recently, been largely ignored in the human
factors cognitive engineering / decision-making area. This is surprising, as extensive empirical evidence
indicates that emotions, and personality traits, influence human perception and decision-making. This is
particularly the case in crisis situations, when extreme affective states may arise (e.g., anxiety). The
development of more complete and realistic theories of human perception and decision-making, and
associated computational models, will require the inclusion of personality and affective considerations.
In this paper, we propose an augmented version of the recognition-primed decision-making theory,
which takes into consideration trait and state effects on decision-making. We describe a cognitive
architecture that implements this theory, and a generic methodology for modeling trait and state effects
within this architecture. Following an initial prototype demonstration, the full architecture is currently
being implemented in the context of a military peacekeeping scenario.

INTRODUCTION stimuli, and a bias towards negative appraisal


Both stable characteristics (traits) and transient (Matthews et al., 2000). Low emotional stability is also
emotions and moods (states) influence behavior. (This associated with increased likelihood of negative affect
is, of course, in addition, to cognitive ability and punishment-avoiding behavior, while high
differences, skill level and training, individual histories, extraversion is associated with increased likelihood of
and contextual factors.) Much empirical and some positive affect and reward-seeking behavior (Matthews
theoretical research exists regarding the nature of these and Deary, 1998).
influences on perceptual, cognitive, and motor Traits also contribute to the dynamic characteristics
processes (e.g., Matthews et al., 2000; Matthews & of affective state generation and expression, which is
Deary, 1998; Williams et al., 1997; LeDoux, 1989; one of the key mechanisms through which traits exert
Ekman and Davidson, 1994). These influences exist their influence. In other words, particular trait value
both at the “lower” levels of processing (e.g., attention combinations map onto specific values of affective state
orientation during an acute fear episode, increased trigger thresholds, growth and decay rates, and
working memory capacity correlated with positive intensity levels. Thus, for example, low emotional
affect), and at “higher” levels involving goals, situation stability correlates with lower trigger thresholds, steeper
assessments, expectations, and self schemas (e.g., growth, slower decay and higher intensity for fear and
complex feedback relationships between affective state anxiety states, as well as more generalized expressions
and self-schemas (Matthews et al., 2000a)). of anxiety (Rothbart, 1994; Matthews and Deary, 1998).
As might be expected, traits tend to exert their However, attempts to operationalize the nature of
influence via more stable structures (e.g., types of these influences at a sufficient level of detail to allow
schemas stored in long-term memory (LTM), computational modeling have only recently begun to
preferential processing pathways among hc tiona l emerge. Such models are particularly relevant to
components), whereas states tend to produce transient cognitive engineering and decision-making research
changes that influence the dynamic characterisiics of a undertaken within human factors, since they promise to
particular cognitive or perceptual process (e.g., attention produce computational models of cognition and
and WM capacity, speed, and accuracy). For example, performance that more closely resemble the human user
low emotional stability (also referred to as neuroticism) 1 operator.
is correlated with a predominance of negative, threat- While much progress has been made in developing,
and self-related schemas in long term memory, as well refining and evaluating theories of situation assessment
as schemas pertaining to affect (Matthews et al., 2000). and decision-making (e.g., Klein, 1997; Endsley, 2000)
It is also correlated with attentional and situation and some studies focus on individual differences
assessment preference for self-related and affective (Gugerty and T h e , 2000) and stress (e.g., Cannon-
304 PROCEEDINGS of the HUMAN FACTORS AND ERGONOMICS SOCIETY 46th ANNUAL MEETING -- 2002
Proceedings of the Human Factors and Ergonomics Society 46th Annual Meeting -- 2002

Bowers and Salas, 1998), little attention has been paid cognitive arihitecture described below. The
to explicitly incorporating affective and personality requirements included the following:
influences within these theories, at a level that would 0 Explicit representation of specific processes within
enable computational modeling. A notable exception is dedicated architecture modules (e.g., attention,
the work of Matthews (e.g., Matthews et al., 2000b). situation assessment, expectation generation, affect
Ignoring these factors leads to incomplete or appraisal and goal management).
erroneous task analysis. The resulting system and user 0 Explicit representation of a number of internal mental

interface designs may then lack the necessary constructs: cues, situations, expectations and goals;
accommodations to these pervasive trait and state with goals representing the primary means of
effects on cognition and behavior, with consequent risks coordinating perception and action selection, and
of suboptimal or inappropriate user / operator mediating the influences of states and traits.
performance. 0 Explicit representation of the mechanisms of
In this paper, we describe an augmented theory of reciprocal influences among mental constructs and
recognition-primed decision-making model (RPD) the user / operators’s affective states.
(Klein, 1989), that takes into account the effects of 0 Dedicated module for deriving the user / operator’s
personality traits and emotion states. The theory is affective state, resulting from the current situational
embedded within a symbolic cognitive architecture, and internal operating context.
which is parameterized to support the modeling of a 0 Distinction between task and self constructs, to
broad range of individual differences, with emphasis on accommodate differences in performance focus,
traits and states. The architecture explicitly represents resulting from differences in trait / state profiles,
processes and structures necessary to model the effects and uniform processing of these constructs via the
of these factors on attention, situation assessment and same underlying mechanisms, which supports
action selection. The model is currently being evaluated smooth shifts in focus between task- and self-
in the context of a Stability and Support Operation relevant perception, processing and performance.
(SASO) military peacekeeping scenario.
DESCRIPTION
METHOD We briefly outline the modeling methodology and the
Two key issues need to be addressed in developing architecture below.
the affect-augmented version of the RPD theory, and an
associated computational model: Individual Differences Modeling Methodology
0 What is the best means of modeling trait and state The core component of our approach is a generic
effects within a cognitive architecture? methodology for modeling individual differences within
0 What additional processes and mental constructs symbolic cognitive architectures, via parametric
must be introduced to the existing RPD theory, to manipulations of the architectural processes and
enable the modeling of trait and state effects? structures. A key objective is to provide flexibility
To address the first issue, we developed a generic regarding the ypes of factors selected, and the nature of
methodology for representing the effects of a broad their influence on perception and cognition. This
range of individual differences factors, including traits approach is described in more detail elsewhere
and states, within a cognitive architecture: the (Hudlicka, 2002b; Hudlicka & Pfautz, 2002a; Hudlicka,
Methodology for Analysis and Modeling of Individual 1997; also appears in Pew & Mavor, 1998, pp. 261-
Differences (MAMID). The underlying assumption of 267). The distinct individual differences factors are
the approach is that the combined effects of individual mapped onto distinct configurations a simulated agent’s
differences can be modeled by varying the cognitive cognitive architecture parameters, which in turn
architecture parameters that control both the processing produce different processing within the architecture,
and the structure of the architecture and its constituent and ultimately lead to different behavior.
modules and knowledge-bases. The effects of individual differences factors can be
To address the second issue, we analyzed existing modeled in two ways: (1) by manipulating the
empirical evidence and theories regarding the effects of architecture processing parameters (e.g., working
traits and states on perception, cognition, and behavior, memory and attention capacity and accuracy), to reflect
and the associated mechanisms mediating these effects. a particular personality trait (e.g., preference for self-
This analysis led to the identification of processing and vs. task-relevant cues), or particular emotion (e.g., bias
structural requirements that motivated the design of the towards detection of threatening cues), thereby biasing
PROCEEDINGS of the HUMAN FACTORS AND ERGONOMICS SOCIETY 46th ANNUAL MEETING -- 2002 305
Proceedings of the Human Factors and Ergonomics Society 46th Annual Meeting -- 2002

the behavior in a particular direction; and by (2) constructs was necessary to demonstrate the trait-
manipulating the architecture content (e.g., structure influenced differences in self vs. non-self focus.
and content of the long term memory stores), to reflect These considerations led to the design of a symbolic,
the desired characteristics (e.g., stable trait-based integrated cognitive architecture implementing an
perceptual and decision-making biases, etc.). Figure 1 affect-augmented recognition-primed decision-making
provides an illustration of the general relationship model. The architecture consists of the following
between the traits / states, the architecture parameters, modules: sensory pre-processing, which translates the
and the architecture. Table 1 lists examples of cognitive incoming raw data into high-level task-relevant
architecture parameters. perceptual cues; attention, which filters the incoming
cues and selects a subset for M e r processing;
situation assessment, which integrates individual cues
into an overall situation assessment; expectation
generation, which projects the current situation into one
or more possible future states; aflect appraisal,, which
derives the affective state fiom both static factors (traits,
individual history) and dynamic factors (current
affective state, current situation, goal, expectation); goal
selection, which selects the most relevant goal for
achievement; and action selection, which selects the
most suitable action for achieving the current goal
within the current context.
Figure 1: Schematic Illustration of MAMID Consistent with the RPD model, the processing
Trait / State Modeling Approach and Architecture emphasis is on the perception of relevant cues, their
integration into a task-relevant situation, and the
Table 1: Examples of Architecture Parameters selection of an appropriate action by mapping the
Available for Modeling Individual Differences situation onto an action within the available behavior
repertoire. Figure 2 illustrates the cognitive
architecture, its constituent modules, and the mental
constructs that comprise the input and output of these
modules; (i.e., cues, situations, expectations, goals, and
actions). The early version of the MAMID cognitive
architecture was loosely based on the SAMPLE
architecture (Zacharias et al., 1995; Harper &
Zacharias, 2002), which also implements the RPD
model, but places more emphasis on the early
perceptual filtering and does not include expectations
and goals, or affective processing and the effects of
traits and states.
Cognitive Architecture
Architecture Parameter Space
The structure of the architecture was motivated in The effects of individual differences are modeled by
part by cognitive theories regarding attention, memory, translating distinct trait / state profiles (e.g., high
decision-making; by the existing conceptual structure of extraversion, low neuroticism, low anxiety, etc.) into
the RPD model; by empirical evidence .regarding
specific values of the architecture parameters. These
trait/state effects; and by theories of the mechanisms then determine the speed and capacity limitations of the
mediating these effects. individual modules, and control the ranking, selection,
Thus, an explicit attention module was necessary to and processing speed of the individual constructs within
demonstrate the trait / state induced attentional biases; the modules.
goals were necessary to demonstrate the influence of
goal-situation mismatch on affect generation;
expectations were necessary to demonstrate effects of
expectation valence on affect and action selection; and
a distinction between self- and task-relevant mental
306 PROCEEDINGS of the HUMAN FACTORS AND ERGONOMICS SOCIETY 46th ANNUAL MEETING -- 2002
Proceedings of the Human Factors and Ergonomics Society 46th Annual Meeting -- 2002

threatening, in part because of the highly-ranked high-


threat cues and lack of balancing and contextual cues
providing evidence to the contrary, and in part due to
trait-associated predominance of threatening schemas in
LTM (content bias), and state-associated preference for
recalling such threatening schemas (process bias).
This interpretation will further increase the highly
anxious commander’s state of anxiety, causing him to
select a self-relevant goal to reduce anxiety, and a task-
relevant goal to defend against attacking enemy. These
goals will then cause the high-anxious commander to
select inappropriate or non-optimal actions. First,
addressing the ‘reduce-anxiety’ goal, the commander
will attempt to reduce anxiety by excessive, irrelevant
communication behavior, not warranted by the current
situation. Second, addressing the ‘defend-from-danger’
goal, the commander will fire into the crowd, also not
an appropriate action given the relative size and combat
Figure 2: MAMID Cognitive Architecture, Mental effectiveness of an armored unit versus a crowd.
Constructs and Emotions that Comprise the Input and In contrast, the low-anxious commander will not be
Output of the Architecture Modules subject to these trait and state biases, either by the
This affect-augmented RPD model is thus able to process-based focus on threatening cues, situations, and
account for behavior which might otherwise appear self and danger related goals, or by the content of the
‘irrational’ (e.g., shift in focus away from the task when corresponding LTM structures (e.g., threat and danger
a crisis situation is occurring), by modeling the explicit schemas, self schemas, and anxiety-reduction and self-
threat-induced shift to a self-relevant goal of anxiety- defense activity repertoire). Figure 3 provides a
reduction versus a task-relevant goal to continue the summary of the effects of these distinct profiles for each
mission. of the cognitive architecture modules.
FINDINGS
Following an initial prototype demonstration
(Hudlicka and Billingsley, 1999), the full architecture is
currently being implemented and nearing completion.
The current demonstration scenario is within a military
context, where multiple commanders are defined in
terms of distinct individual differences profiles, and
their behavior is modeled by instances of the cognitive
architecture. Alternative mission outcomes can then be
observed, resulting from individual commander’s
distinct reactions to the same set of events (e.g., Figure 3: Distinct Commander Profdes and Behaviors
encountering a hostile crowd en route to an objective). DISCUSSION
For example, when two distinct commander types We have operationalized the conceptual RPD model
(e.g., ‘low anxious’ and ‘high anxious’) encounter the within a computational cognitive architecture and
same situation (e.g., a hostile crowd blocking their augmented the existing model structure with a dedicated
route), their behavior differs significantly. This is due to affect appraisal module and explicit representations of
a variety of state / trait effects across the spectrum of trait and state effects on perfonnance. Preliminary data
the decision-making process. from the initial MAMID prototype, and the on-going
During attentional cue filtering, an anxious full-implementation, demonstrate distinct differences in
commander will focus on threatening cues and neglect individual commander behavior, as a fimction of their
or downgrade other cues. This will result in highly- individual profiles, and consequent significant
ranked cues regarding the crowd attack reaching the differences in the final mission outcome. The modeling
situation assessment module. There, the anxious environment allows the analyst to trace the specific trait
commander will interpret the crowd attack as highly
PROCEEDINGS of the HUMAN FACTORS AND ERGONOMICS SOCIETY 46th ANNUAL MEETING -- 2002 307
Proceedings of the Human Factors and Ergonomics Society 46th Annual Meeting -- 2002

/ state effects through the decision-making sequence, ACKNOWLEDGMENTS


starting with incoming cues and ending with the final The research described in this paper was supported in part by US
Army Contract DASWOI-00-C-3000. W e would like to
outgoing action selection. The environment also acknowledge continued support of the COTR Dr.Bob Witmer, and
supports flexible manipulation of a variety of factors the contributions of the following individuals: Dr.Greg Zacharias of
and parameters, to fine-tune the model performance. Charles River Analytics provided valuable discussions and
Model validation will be performed once the full- suggestions regarding the simulation algorithms, Prof. William
implementation is complete. Both a heuristic Revelle of Northwestern University made suggestions regarding the
individual differences factor selection, and Mr. Ted Fichtl of The
evaluation by subject matter experts (psychologists and Compass Foundation was the subject matter expert. Mark Turnbull
military commanders), and an in-depth comparison with of Contented Software developed the world simulation module. We
empirical data are planned. However, due to lack of also acknowledge the efforts of the Psychometrix software team:
sufficiently detailed empirical data, some of this Lisa Buomano, Jim Helms, Craig Ganoe
validation process may be limited to input / output REFERENCES
validation. Cannon-Bowers, J. A. & Salas, E. (1998). Decision Making Under Stress.
One of the key contributions of a computational Washington, DC: APA.
Ekman, P. & Davidson, R.J. (1994). The Nature of Emotion. Oxford..
approach is the ability to generate hypotheses for Endsley, M. R. (2000). Theoretical Underpinnings of Situation Awareness: A
targeted empirical studies. A tightly-coupled Critical Review. In Situation Awareness Analysis and Measurement,
theoretical-modelingand applied-empirical program of M.R. Endsley & D.J.Garland, eds. Mahwah: NJ: LEA.
Endsley, M. R. (1995). Measurement of Situation Awareness in Dynamic
research promises to produce detailed characterizations Systems. Human Factors, 37(1).
of the effects of traits and states on performance, and Gugerty, L.J. & Tim, W.C. (2000). Individual Differences in Situation
the associated mediating mechanisms. Awareness. In Situation Awareness Analysis and Measurement, M.R.
Endsley and D.J. Garland, eds.Mahwah: NJ: LEA.
Harper, K. & Zacharias, G. (2002). Modeling Attention Allocation and
CONCLUSIONS Multitasking in Computational Human Behavior Representation. In
Proceedings of the IShConference CGF-BR,May, 59-66.
The ability to explicitly model effects of traits and Hudlicka, E. & Pfautz, J. (2002a). Architecture and Representation
states on performance will contribute to the Requirements for Modeling Effects of Behavior Moderators. In
development of more realistic human performance Proceedings of the 11" Conference on CGF-BR,May, 9-20.
Hudlicka, E. (2002b). This time with feeling: Integrated Model of Trait and
models. Such realism is particularly necessary as a State Effects on Cognition arid Behavior. Applied AI, 16:1-31.
variety of 'mission-critical' decision aids and Hudlicka, E. (1997). Modeling Behavior Moderators in Military HBR
computerized systems become increasingly ubiquitous. Models. TR # 9716. Lincoln, MA: Psychometrix Associates, Inc.
(sections appear in Pew and Mavor, 1998)
To ensure effective human-machine interaction, Hudlicka, E. and Billingsley, J. (1999). Representing Behavior Moderators in
across a variety of tasks and populations, cognitive Military Human Performance Models In Proceedings of the 8 '
Conference on CGF-BR. Orlando, FL.
engineers must take into account the broad range of Klein, G. A. (1997). The recognition-primed decision (RPD) model: Looking
individual differences effects on performance. back, looking forward. In C, Zsambok & G. Klein (Eds.), Naturlistic
Existing results, and future developments, of the decision making. Mahwah,NJ: Erlbaum.
Klein, G. A. (1989). Recognition-Primed Decisions. Advances in Man-
MAMID cognitive architecture and modeling Machine Systems Research, 5,47-92.
methodology are relevant from both a theoretical and LeDoux, J.E. (1989). Cognitive-Emotional Interactions in the Brain.
from an applied perspective. The ability to model and Cognition and Emotion, 3(4), 267-289.
Matthews, G., Denybeny, D., & Siegle, G.J. (2000a) Personality .and
explore alternative theories regarding mechanisms emotion: Cognitive science perspectives. In S.E. Hampson (Ed.),
mediating trait and state effects is valuable for basic Advances in personality psychology (Vol. 1). London: Routledge.
research and understanding of these pervasive effects, Matthews, G., Davies, D.R., Westerman, S.J. and Stammers, R.B. (2000b).
Human Performances Cognition, Stress, and individual differences.
both positive and negative, on performance. The ability Philadelphia, PA: Taylor and Francis.
to model distinct operator / user types, based on a Matthews, G., and Deary, 1.J. (1998). Personality Traits. Cambridge, U K :
Cambridge.
definition of their individual differences profile, has Pew, R.W. and Mavor, A.S. (1998). Representing Human Behavior in
applications in a variety of areas of interest to the Military Simulations. Washington, DC.National Academy Press.
cognitive engineer, including user interface design and Williams, J.M.G., Watts, F.N., MacLeod, C., and Mathews, A. (1997).
Cognitive Psychology and Emotional Disorders. NY: John Wiley.
task allocation; individual and team training Zacharias, G.L., Miao, A.X, Illgen, C., and Yara, J.M. SAMPLE: Situation
environments; performance analysis and assessment; Awareness Model for Pilot-in-the-Loop Evaluation. In Proceedings of
the Ist Annual Conference on Situation Awareness in the Tactical Air
and behavior modeling, simulation, and prediction. Environment. Wright-Patterson AFB, OH: CSERIAC.

Das könnte Ihnen auch gefallen