affective computing intro class
TRANSCRIPT
8/26/11
1
Affective Computing Intro Class"
!!
Stacy Marsella & Jon Gratch!!!Institute for Creative Technologies!University of Southern California!
2!
Background"
Computational Models of Human Behavior"!Cognitive and Emotional processes"
!Social Interaction & Theory of Mind!
!Nonverbal behavior (Facial Expressions, gesture, posture, gaze…)!Applications!
!Virtual Humans, Social Simulation, Interactive Drama"Architectures!
!Tools: PsychSim/Thespian, SmartBody, NVBG!
8/26/11
2
Outline"
! Emotion and it’s impact on behavior"
! Why model emotion computationally?"– Applications!
! How we model it"– EMA ! !
! Can the model predict human behavior in a game"
Why study emotion?"
�Reason is, and ought to be, only the slave of the passions�
David Hume, 1711-1776
8/26/11
3
""
Marsella & Gratch, Journal of Cognitive Systems Research 2009
What is emotion?"
! Functions associated with individual autonomy"– Rapid, continually adjusting assessment of significant events!– Interruption of behaviors and changing of goals!– Action preparation: energizes body, changes physical orientation!
! Functions associated with social interaction"– Signaling: broadcast information about mental state!– Coordination: orient and coordinate group response!
! Immersion and contagion"
8/26/11
4
7!
Why Study Emotion? Emotion influences cognitive processes"
! Emotions change perception & decision-making"– Angry people blame others/outgroups
(Keltner et al 93; Mackie et al 00) !– Quicker to perceive threats (DeSteno et
al 2000/2004)!
– Underestimate risk (Lerner & Keltner 2000/2001)!
! Emotions impact memory"– Mood Congruent Recall & Learning
(Bower, 91)!
8!
Why Study Emotion? Emotion influences human behavior"
! Emotions change behavior"– Anger primes aggressive responses
(Keltner & Haidt 1999)!– Characteristic displays (Spoor&Kelly 04,
Parkinson 01, Darwin, Ekman)!
8/26/11
5
9!
Why Study Emotion? Emotion influences social behavior"
! Emotions impact social interaction"– Distress elicits helping (Eisenberg et al 89)!
– Anger elicits fear (Dimberg&Ohman96) !
– Negotiators concede more to angry partner (van Kleef et al. 2007)!
– Emotion communicates information to other social actors (Darwin; Parkinson01)!
! Emotions impact social systems"– Improves group utility (Darwin; Frank, 1988)!
Why model emotion? Claim: emotion research has important benefits:"
! Model human behavior"– Inform theory development (Marsella et al. 2010)!
! Inform general theories of intelligence"– Provide insights to expanding general theories !– Simon,67; Minsky, 85; Scheutz, & Sloman, 2001!
! Applications"
8/26/11
6
Applications: Better predictive models of human behavior"! More accurate simulations of
human behavior "! Use: emergency response
preparation and disaster planning"
! How:"– Model Emotion’s impact on decision-
making!– Model Emotional contagion!
Applications: Enhance human computer interaction"
! Recognize and respond to human�s emotions"– Intelligent Tutoring Systems!
! (Lisetti & Schiano, 2000; Conati & MacLaren , 2004)!
– Security (Ekman & Friesen)!
! How:"– Model Player, User, Student !– Recognize emotional
behavior!! Facial Expression Recognition
(Movellen, Cohn)!!
! )!
8/26/11
7
Applications: Enhance human computer interaction II"
!
! Elicit emotional responses to engage user"
! Design-A-Plant’s Herman the Bug"
14!
Applications: Virtual Humans"
Intelligent agent that supports social interactions with human users in virtual reality!
!
Characters with a brain !– Reason about environment!– Understand and express emotion!– Communicate through speech & gesture!– Play the role of teachers, peers, adversaries!
8/26/11
8
15!
Support range of applications""Social-skills training"
Interview skills (TacQ/IEWTPT)!Meeting and negotiation skills (ELECT BiLAT)!Command and personal issues ( FITE, VOLT)!
"Clinical training"Medical or psychiatric assessment (Virtual Patient; People factory)!Medical triage (Sick Call)!
"Persuasive interfaces"Health seeking / social support (Sim Coach, TOPPS-VW)!Medical adherence / outpatient procedures (Bickmore)!Tutoring (Lester; Graesser; !Advertising / Marketing / Customer support (CodeBaby)!Informal science education (Twins)!
"Assessment and Experimentation"Explore factors that promote rapport / self-disclosure (Rapport Agent)!Personality assessment (Lester)!
"Entertainment"Gunslinger!!!!
Health Interventions • Social Skills Training (Marsella et al) • Autism (Tartaro & Cassell)
Diagnosis & Assessment • ADHD Assessment (Rizzo et. al)
Education and Training • Intelligent tutoring systems (Lester et al; Graesser) • Health care provider training (Lok et al)
Increased interest in applications"
Entertainment • Façade (Mateas and Stern) • Thespian (Si et al.) !
Informal Science Education • Museum Guides (Boston Science Museum, Cooper-Hewett)
8/26/11
9
Health Interventions • Bullying (Aylett et al) • Coping (Marsella et al) • Autism (Tartaro & Cassell) • AIDS Prevention (Miller et al.)
Virtual Patients to train health care providers • Kron et al., Lok et al.; Kenny et al
Social Skills Training (& Entertainment) w/ Virtual Humans"
Cross-cultural trainer for Military • Language (Johnson et al) • Negotiation (Kim et al)
Medical CyberWorld�s Robin
Gunslinger
Entertainment • Gunslinger
Common Theme: Central Role of Emotion • User immersed in emotional scenario
• Informing a patient they have a fatal disease • Confronting a gunfighter
• Virtual human must convey emotion • Patient finding out she has a fatal disease • Bartender having a gun pointed at him
8/26/11
10
Cognition
Understanding
Behavior Perception
Speech Prosody
Language
Dialogue
Generation
Planning
Annotated Text
Intent
Annotated Text
Emotion
Nonverbal
Gesture
Language Nonverbal
Speech
Gesture
Intent
Rea
ctiv
ity
Research &Technology of
Virtual Humans"
Perception • Morency NLP • DeVault, Hovy, Traum Emotion/Cognition • Gratch, Marsella, Pynadath Behavior Generation • Lee & Marsella Animation • Chiu, Marsella & Shapiro
EMA: Computational Model of Emotion"
! How do we computationally model emotion?!– Both the causes and cognitive/behavioral
consequences of emotion?!
! How do we evaluate/validate these models?!– Can it predict human behavior?!– And model users?!
! How do we use the model in virtual humans?!
8/26/11
11
Theoretical Perspective:�Appraisal Theory
Magda Arnold
• Emphasizes cognitive antecedents of emotion – Emotion arises from an evolving subjective
interpretation of person’s relation to their environment – Well-suited to computational realization
• Emotion arises from inference over representations
(Arnold, Lazarus, Frijda, Scherer, Ortony et al.)
22!
(act on world) (act on self)
Emotion
Coping Strategy
Action Tendencies “Affect” Physiological
Response
Problem-Focused Emotion-Focused
Environment Goals/Beliefs/ Intentions
Theoretical Framework: Appraisal Theory " " "(Arnold, Lazarus, Frijda, Scherer, Ortony et al.)"
Desirability
Expectedness
Controllability
Causal Attribution
Appraisal
8/26/11
12
23!
(act on world) (act on self)
Emotion
Coping Strategy
Action Tendencies “Affect” Physiological
Response
Problem-Focused Emotion-Focused
Environment Goals/Beliefs/ Intentions
Theoretical Framework: Appraisal Theory " " " " (Arnold, Lazarus, Frijda, Scherer, Ortony et al.)"
Desirability
Expectedness
Controllability
Causal Attribution
Resignation
Distancing
Wishful Thinking
Take action
Seek support
24!
Appraisal Theory as Architectural Specification"
To derive a computational model need to model:"• Representation of Person-Environment relation!• Derivation of appraisal values.!• Map from appraisals to emotions.!• Behavioral, cognitive consequents of emotion!
– Coping!
• EMA: EMotion & Adaptation!– (Gratch & Marsella 2005; Marsella & Gratch 2009)!
8/26/11
13
25!
Computational Appraisal Models"
Appraisal Theories
AR Elliott
EM Neal Reilly
FLAME El Nasr
EMILE Gratch
CBI Marsella
EMA Gratch/Marsella
Fatima Dias
PEACTIDM Marinier
ALMA Gebhard
WASABI Becker-Asano
ACRES Swagerman
WILL Moffat Frijda!
OCC!
Lazarus!
Scherer!
ParleE Bui
THESPIAN Si et al.
TABASCO Staller&Petta
ActAffAct Rank
EMA: EMotion & Adaptation!(Gratch & Marsella 2005; Marsella & Gratch 2009)!
Planning Perception Dialogue Action
Causal Interpretation Working memory of plans, beliefs, desires, intentions
Cognitive Operations (inference)
Past Events
Future Plans Present
EMA: Person-Environment Relation
Future Act Cause: self Intend: yes Probability: 50%
Goal Utility: 50 Probability: 50% Intend-that: True
Goal Utility: 50 Probability: 100% Belief: False Past Act
Cause: Other Intend: yes Prob: 100%
Facilitates Inhibits
8/26/11
14
EMA: Appraisal Derivation Model"
• Appraisal as evaluation of the causal interpretation"• Define appraisal variables in terms of features!• Fast, parallel pattern matching!
• Appraisal Variables"• Desirability: Does proposition facilitate/inhibit agent’s utility!• Likelihood: Likelihood of state.!• Expectedness: Could truth value of a state be predicted.!• Controllability: Can the outcome be altered by agent!• Changeability: Can outcome be altered by other causal agent!• Causal Attribution: what agent is blameworthy/praiseworthy!
• Scope"• Blueprint for a social agent!
EMA: Coping Model"
• Sequential, deliberate, mediated by focus of attention"• Problem-focused � Take Action, Make Plans"• Emotion-focused � Alters Attention, Beliefs & Desires:"
• Attention!• Avoidance � Take action that alters attention!
• Beliefs!• Wishful Thinking � Change belief / likelihood!• Shift blame � Change causal attribution!
• Desires!• Distancing � Lower utility of desired, threatened state!• Find silver lining � Change utilities!
• Intentions!• Resignation � Drop intention to achieve a desired state!
• Key: Finding ecological niche"• Concretize Coping Theory: Framework for modeling emotion’s impact!
8/26/11
15
Future Act Cause: self Intend: yes Probability: 50%
Goal Utility: 50 Probability: 50% Intend-that: True
➧ Past Present Future ➨
Goal Utility: 50 Probability: 100% Belief: False
Past Act Cause: Other Intend: yes Prob: 100%
Challenge Desirability: 50 Likelihood: 50% Causal Attribution: self Control: Moderate Emotion: Hope(25)
Threat Desirability: -50 Likelihood: 100% Causal Attribution: Other Control: moderate Emotion: Anger(50)
Facilitates
Inhibits
Appraisal
Threat Desirability: -100 Likelihood: 50% Causal Attribution: Other Control: Low Emotion: Sadness(50)
➧ Past Present Future ➨
Goal Utility: 50 Probability: 100% Belief: False
Past Act Cause: Other Intend: yes Prob: 100%
Threat Desirability: -50 Likelihood: 100% Causal Attribution: Other Control: moderate Emotion: Anger(50)
Inhibits
Coping
Challenge Desirability: 50 Likelihood: 50% Causal Attribution: self Control: Moderate Emotion: Hope(25)
Future Act Cause: self Intend: yes Probability: 50%
Goal Utility: 50 Probability: 50% Intend-that: True
Facilitates
Resignation (abandon goal)
8/26/11
16
Surprise Fear Anger
Empathy Humor/Relief
Emotional Dynamics
Compassion
Relevance
Implication
Coping Potential
Normative Significance
Causal Interpretation (“Working Memory”) Environment
Control Signals
Appraisal Frames
Affective State
Appraisal
Coping
Inference Action
Appraisal
Scherer’s Sequential Checking Model EMA
Inference
Dynamics in perceived world relationship
Environment
Dynamics in the world
Action
Dynamics through action
Marsella & Gratch JCSR 2009
8/26/11
17
Bird Flies In Personal Health Desire: Self (100)
Satisfied: True Probability: 80%
Surprise
Perspective: Self Unexpectedness: High
Controllability: Low Blame/Credit: unresolved
Appraisal
Virtual Actor’s Appraisal of Bird Flying in Window
Seek Info
Coping: Seek More Information
Cop
ing
Bird Attack Responsibility: Bird
Bird Flies In Personal Health Desire: Self (100) Satisfied: False Probability: 60%
Fear: 48
Further Inference: The Bird will attack
Perspective: Self Desirability: Undesirable
Controllability: Low Blame/Credit: Bird
Appraisal
Threatens
Move Away
Avoid Threat C
opin
g
Coping: Avoid
8/26/11
18
Bird Attacking Responsibility: Bird
Bird Flies In Personal Health Desire: Self (100) Satisfied: False Probability: 50%
Anger: 48
Re-Appraisal & Further Planning
Perspective: Self Desirability: Undesirable Controllability: Medium
Blame/Credit: Bird
Appraisal
Threatens
Attack Threat
Whack Bird
Coping: Plan
Cop
ing
Threatens
Bird at safer dx
Marsella & Gratch JCSR 09
Control Signals
Causal Interpretation (Goals, Beliefs, Causal Relations, Plans, Intentions) Environment
Appraisal Frames
Affective State
Appraisal
Coping
Plan Inference
Belief Formation
Explanation
Action
Dialogue
Emotion as Meta-control system
8/26/11
19
37!
Validity & Application"
• Does the model predict human emotional reactions and their impact on behavior?"
"!!
Theoretical Framework: Computational Appraisal Theory"
(act on world) (act on self)
Emotion
Coping Strategy
Action Tendencies �Affect� Physiological
Response
Problem-Focused Emotion-Focused
Environment Goals/Beliefs/ Intentions
Desirability
Expectedness
Controlability
Causal Attribution
Appraisal
! Autonomy: What are emotion�s cognitive consequences?"! Multiagents: What are emotion�s social consequences?"
8/26/11
20
Cognitive & Behavioral Consequences: Coping"
! Human tendency to alter mental beliefs and commitments in response to emotion "
! Often characterized as irrational"! But argued to serve important adaptive functions""Simon 1967, Frank 1988, Damasio 1994, Mele 2001"
! Questions:"– Can EMA predict such �distortions�!– How do they differ from �rational� models!!
Danger!
Danger!
Play%Game%Lose%
Win%
p=.8
p=.2 Utility= 20
Intention(play) EU=4%
! Rational models decouple preferences and beliefs"– Desires shouldn�t change beliefs (and vice versa)!
! e.g., Just wanting something shouldn�t make it true!– Preferences fixed over time!
8/26/11
21
Wishful Thinking
Play%Game%Lose%
Win%
p=.8
p=.2 Utility= 20
Intention(play) Sad%
Hope%Joy%
Distancing
Fear%
Utility= 10 p=.6
p=.4
Resignation
• Coping serves to �confound� beliefs and desires"! Emotion-biases on decision making (Loewenstein & Lerner, 2003)!! Cognitive dissonance (Festinger57)!! Motivated inference (Kunda87)!
– Few attempts to model computationally (Marsella&Gratch; Dias)!
42!
Coping Study (Marsella, Gratch, Wang, Stankovic, ACII09)"
Can EMA predict human behavior in experimental games" !!
! ! ! ! ! Does success/failure during game alter!! ! ! ! ! !Desire to win?!! ! ! ! ! !Intention to play?!! ! ! ! ! !Perceived likelihood of winning?!
!! ! ! ! ! !!
""Model Driven Experimentation Paradigm"– Use EMA to generate task specific predictions!– Evaluate predictions through human subject experiments!
"
8/26/11
22
43!
Build Model"Simulate some games"! �Winning� vs. �Losing�!! Developed and validated
in 2 pilot studies!
Note EMA predictions"! Automatically derives
emotion and coping tendencies!
! Appraisals and coping tendencies constitute a set of predictions that can be tested against data!
EMA Predictions"
44!
EMA�s Coping Predictions (subset)"
! Beliefs (wishful thinking)"– H3: Losing interacts with utility to predict probability bias!
! Players that want to win will perceive higher win probability in lose condition!
Emotion will �irrationally� change"! Desires (distancing)"
– H1: Subjective utility of winning will drop as player loses!! H1a: Magnitude of subjective utility predicts intensity of effect!
! Intentions (resignation)"– H2: Intention to play will drop as player loses!
! H1a: Magnitude of subjective utility predicts intensity of effect!!
8/26/11
23
Con
fede
rate
Hid
den
Cam
era
Par
ticip
ant
Prior Expectations LOSING LOST GAME
Assess emotions and coping
Time 0 Time 1 Time 2
Prior Expectations WINNING WON GAME Winning condition
Losing condition
Human subjects study 100 participants (2 conditions)*
Instruments
– Measured Emotions and Appraisals – Self-Report: based on Ellsworth and colleagues� appraisal
questionnaire (e.g., Treynor, Ellsworth, & Gonzalez, in preparation; also see Smith & Ellsworth, 1985)
– Measured coping – Self-Report: Computer Coping Questionnaire
(based on SCPQ; Perrez & Reicherts, 1992)
– Behaviorally: Response time
8/26/11
24
0 5
10 15 20 25 30
0 10 20 30 40 50 60 70 80 90 100 Probability
0
20
40
60
80
100
0 10 20 30 40 50 60 70 80 90 100 Probability
0 5
10 15 20 25 30
0 10 20 30 40 50 60 70 80 90 100 Probability
0
20
40
60
80
100
0 10 20 30 40 50 60 70 80 90 100 Probability
JOY FEAR
HOPE SADNESS
Emotion
Subjects find experience to be emotional
! Emotions change predictably throughout game
! Appraisals of utility and probability predict emotion intensity
Hi win motivation
Lo win motivation Winning Losing Winning Losing
48!
Coping"
! Emotion distorts desires and intentions
! H1 and H2 supported: losers distance and resign (sour grapes)
winners
losers
�Rational� behavior
Self-reported desire to win Self-reported willingness to play
losers
! But also �coping� with winning observed ! EMA implementation makes no prediction
winners
! Rationality assumption: desire should be independent of probability
8/26/11
25
49!
Individual differences: Impact depends on initial preferences"
! H1a and H2a also supported: strength of an individual�s desire predicts strength of coping response ! In Lose condition: Subject�s with initial high desire to win distanced
and resigned more
! In Win condition: Subject�s with initial low initial desire to win became more engaged (EMA makes no prediction)
50!
Discussion"
! Computational appraisal models (e.g., EMA) can improve predictive power of models"
! Overall, results suggest that coping moves subjects toward more positive emotional states"– Distancing and resignation of losing, high motivated subjects reduces
negative emotionality of threat to high utility goal!– Increased engagement of winning, low motivated subjects enhances
positive emotions from previously unattainable goal!
! Suggests appraisal theory as a tool for understanding and influencing user’s desires and intentions"
8/26/11
26
Behavioral Consequences: social influence"
! If an agent expressed its emotions, would you care?"
! Fact: People influenced by partner�s expressions"– Guilt/Anger fosters cooperation in Prisoner�s Dilemma (Frank)!– Anger promotes concessions in Integrative Bargaining (van Kleef)!– Smiling yields cooperation in Trust Games (Krunhumber)!
! Question: is there value in communicating emotion"– Give agents different goals!
! Cooperative agent values joint reward!! Self-interested agent values individual reward!
– Appraise events in the game (according to EMA)!! Cooperative agent �feels� guilty if it exploits human!! Self-interested agent �feels� satisfaction!
– Display resulting emotions!! Are people influenced by displays of agent emotion?"
!"
Effects of Agent Emotions on People in experimental games"
8/26/11
27
! Two studies""Can emotion promote cooperation?"
– Cooperative vs. unemotional agent!
"Does impact depend on agent goals?"– Cooperative vs. Individualistic emotions!!Create patterns of displays consistent with!!cooperative or individualistic goal orientation!!!
! Play iterated prisoner�s dilemma for chance to win $100"! Participants believe they are playing against computer"! Agent behavior identical (tit-for-tat) except for face expressions"! Predict greater cooperation with cooperative agent""
Studies"
8/26/11
28
! Greater cooperation with emotional agent!
! More cooperation with cooperative emotions!
! Same emotion (e.g., smile) has different impact depending on context in which it is produced!
! Mechanism still unclear!– Is emotion just information?!– Is it inducing emotional
reactions!– Some evidence for the later in
more recent studies!
9.16!
6.80!
0.00!1.00!2.00!3.00!4.00!5.00!6.00!7.00!8.00!9.00!
10.00!
Cooperative! Individualistic!
Aver
age
Num
ber o
f Co
oper
atio
n"
Overall cooperation rate
Cooperation rate by round
Results"
Summary: should emotions influence machine behavior? "! Social science theory can inform social agent design"
– In this talk we focused on emotion!– Use appraisal theory as architectural specification for agent design!
! Value for modeling and predicting human behavior"– EMA enhanced predictive power of classical decision models!!Predicted changes in desires and intentions in competitive game!
! Value for shaping human computer interaction"– Virtual human emotion alters human behavior in experimental games!– Not to mention considerable research into recognizing and reacting to
human emotional cues (Picard; Graesser; Movellan & Bartlett; Pantic; Cohn;…)!
8/26/11
29
Broader picture"
! AI born from interest in human behavior"– The cognitive revolution in psychology, linguistics, !!neuroscience, anthropology and philosophy!
! Emotion important in early discourse "– Simon, H. A. (1967). "Motivational and emotional controls of cognition."
Psychological Review 74: 29-39."
"but fell from favor "– as AI turned to formal models of individual problem solving!
Simon & Newell
! AAMAS a crucible for a social cognitive revolution"– Social cognitive theory clearly can benefit man-machine interaction!– Can it inform general theories of artificial intelligence?!