Depth of Feelings: Modeling Emotions in User Models and Agent Architectures

Depth of Feelings: Alternatives for Modeling Affect in User Models & Cognitive Architectures Eva Hudlicka Psychometrix Associates Blacksburg, US [email_address] psychometrixassociates.com TSD 2006 Masarykova Universita, Brno, Czech Republic September 15, 2006

“ Diseases of the Mind”* Are emotions….. *Immanuel Kant

“ reason is, and ought only to be the slave of the passions” Hume, 1739 Or are emotions essential for adaptive intelligent behavior…

Emotions in “Human” Interaction “ too little…”

Emotions in Human Interaction “ too much..”

Outline Motivation & Objectives Emotions – Background Info Computational Models of Emotion Framework for Model Analysis Summary & Conclusions

Emotions in HCI: State-of-the-art KISMET - Cynthia Breazeal, MIT Media Lab

Emotions in HCI: State-of-the-art Agent Max - Becker-Asano et al.

Requirements for Affective HCI Affective User Model / Cognitive-Affective Architecture Emotion Sensing & Recognition “ Emotion” Expression OR? GRETA, Fiorella de Rosis, U. Bari

Why Include Emotions in User Models & Agent Architectures? Emotion is a critical component of social interaction & individual motivation Affective user models are more realistic, enabling: Socially-appropriate dialogue (speech tone, content, turn taking) More effective dialogue (persuasion, empathy) Infer implied meaning & motivation Predict affective reaction to system’s utterances Affective agent architectures enable: Socially-appropriate responses and behavior May improve agent autonomy in complex, uncertain environments Affect-adaptive user interfaces and responses

Definition(s) of Emotions (See: roles & characteristics of emotions…) Evaluative judgments of world, others, self …in light of agent’s goals and beliefs

Roles of Emotions Intrapsychic Interpersonal WHAT? * Social coordination * Rapid communication of behavioral intent; HOW? Express emotions via: -Facial expression -Speech (content & properties) -Gesture, Posture -Specific actions WHAT? * Motivation * Homeostasis * Adaptive behavior HOW? - Emotion generation (appraisal) Emotion effects (processing biases) Global interrupt system Goal management Prepare for coordinated actions

How Do We Recognize an Emotion if We See One? Manifested across multiple , interacting modalities: Somatic / Physiological (neuroendocrine - e.g., heart rate, GSR) Cognitive / Interpretive (“Nothing is good or bad but thinking makes it so…”) Behavioral / Motivational (action oriented, expressive, ‘visible’) Experiential / Subjective (“that special feeling…”, consciousness) Much terminological confusion can be attributed to a lack of consideration of these multiple modalities of emotions e.g., Is emotion a feeling or a thought? - It’s both

Simple Fear “Signature”: Large, Approaching Object Increased heart-rate; Attacked? Crushed? Flee? Freeze? Feeling of fear Cognitive Subjective

A Taxonomy of Affective Factors Traits Affective Factors NOT ALL TRAITS are affective! Attitudes, Preferences… Affective States Emotions Moods Negative Positive Traits States “ Big 5” … Basic Anger Joy Fear … Complex Shame Guilt Pride …

Core Processes of Emotions Effects of Emotions (on cognition & behavior) Generation of Emotions (via cognitive appraisal) Cognitive-Affective Architecture Stimuli Situations Expectations Goals Cognitive Appraisal Emotions

Emotion Generation via Appraisal Stimuli Appraisal Dimensions Recalled Perceived Imagined Appraisal Process Emotions Existing emotions, moods, traits Goals (desires, values, standards) Beliefs, Expectations

Emotion Generation via Appraisal Stimuli Appraisal Dimensions Recalled Perceived Imagined Appraisal Process Emotions Existing emotions, moods, traits Goals (desires, values, standards) Beliefs, Expectations Domain-Independent Appraisal Dimensions Novelty Valence Goal / Need relevance Goal congruence Agency Coping potential Social and self norms and values

Emotion Effects on Cognition Emotion and cognition function as closely-coupled information processing systems Emotions influence fundamental processes mediating high-level cognition: Attention speed and capacity Working memory speed and capacity Long-term memory recall and encoding Influences on processing and contents Transient biases influence processing Long-term biases result in differences in long-term memory content & structure

Examples of Affective Biases Anxiety Narrows attentional focus Reduces working memory capacity Biases towards detection of threatening stimuli Biases towards interpretation of ambiguous stimuli as threatening Promotes self-focus Positive emotions Increase estimates of degree of control Overestimate of likelihood of positive events Promote creative problem-solving Promotes ‘big picture’ thinking - focus on ‘the forest’ Biases can be adaptive or maladaptive, depending on context

“ Thank God! Those blasted crickets have finally stopped!”

Considerations Guiding Model Requirements Why and when to model emotions? Which emotions? Which aspects of emotions? Which affective processes? Which theory? What level of resolution? (Are the data available?) Which architecture? … Knowledge and data requirements …… Representational and reasoning requirements ……… ..Representational and reasoning formalisms and methods

Why and When to Model Emotions? Research Understand how emotions work in biological agents Applied More effective and ‘fun’ human-computer interaction Decision-support Training & tutoring Recommender systems Entertainment More robust agent behavior

Why and When to Model Emotions? Research Understand how emotions work in biological agents Applied More effective and ‘fun’ human-computer interaction Decision-support Training & tutoring Recommender systems Entertainment More robust agent behavior (Breazeal, 2003) (de Rosis, 2003)

Which Emotions and Affective Factors to Model? Model objectives & application influence selection: User models for decision-support systems in stressful settings: fear, anxiety, frustration, surprise, boredom – probably not pride, shame, guilt Synthetic agents for children: happiness, sadness, fear, anger …also pride, shame User models & agents in training and tutoring: happiness, fear/anxiety, frustration, surprise, boredom

A Taxonomy of Affective Factors States Affective States Emotions Moods Basic Complex Negative Positive Anger Joy Fear Shame Guilt Pride Traits Traits Affective Factors “ Big 5” …

But Exactly Which Aspects of Emotions Should We Model? Recall the multiple modalities of emotions: Somatic / Physiological (infrequent) Cognitive / Interpretive (most frequent) Behavioral / Motivational (most frequent) Experiential / Subjective (infeasible?) Cognitive Subjective

Emotion Roles Emotion Generation Emotion Effects on Cognition & Behavior Which Processes to Model? Social Communication - Coordination … . Intrapsychic: - Goal management - Behavior preparation -…… implement

Computational Tasks for Appraisal Models Stimuli Emotion attributes: Complexity of emotion construct * type * intensity * cause … * direction * … Types of stimuli: Internal / External Real / Imagined Past / Present / Future Domain specific / Abstract appraisal dimensions Complexity of stimulus structure Mental constructs required (e.g., goals, expectations) Stimuli-to-emotion mappings Intensity calculation Nature of mapping process: * Stages & functions * Degree of variability Integrating multiple emotions Emotion dynamics over time Emotions

Most Influential Appraisal Theories in Computational Models Ortony, Clore and Collins (OCC) (1988) Leventhal and Scherer --> Scherer (1984 …) Arnold  Lazarus  Smith and Kirby (1960 …)

Example #1: OCC Appraisal Model

Valenced Reactions Event-based emotions Attribution emotions Attraction emotions Event Related Appraised wrt goals “ Does this promote world peace?” Acts-by-Agents Related Appraised wrt standards “ Was it appropriate for John to rob the bank?” Object Related Appraised wrt attitudes “ Is this appealing to me?”

Valenced Reactions Event-based emotions happy for, pity, gloating.. joy,distress hope, fear gratitude, anger desirability (pleased / displeased) desirability for other (deserving, liking) likelihood praiseworthiness (approve / disapprove) appealingness (like / dislike) degree of autonomy, expectation deviation familiarity Attribution emotions Attraction emotions Fortunes-of-self emotions Fortunes-of-others emotions Prospect-based emotions Well-being emotions pride, shame, reproach love,hate

Valenced Reactions Event-based emotions Attribution emotions Attraction emotions Fortunes-of-self emotions Fortunes-of-others emotions happy for, pity, gloating.. distress Prospect-based emotions Well-being emotions anger reproach love,hate Desirability = low fear Praiseworthiness = low degree of autonomy = high expectation deviation = high

Example #1: OCC Appraisal Model Developed to provide a “computationally tractable model of emotion” Taxonomy of triggering conditions and emotion types Specification of variables affecting intensity “ Global” (physiological state…) “ Local” (appraisal dimensions…) Many implementations (Elliot, Reilly, Bartneck, Andre, Gratch…)

Example #2: Scherer‘s “Component Process Model”

Coping potential Norms Relevance Appraisal variables Novelty Valence Goal relevance Certainty Urgency Goal congruence Agency Stimuli Implications Coping Norms Emotion

STIMULI Novelty Valence Goal relevance Outcome probability Urgency Goal congruence Agency Coping potential Norms high high v. high low other low low high FEAR

Example #2: Scherer‘s “Component Process Model” Emphasis on domain-independent appraisal dimensions (emotion components) Emotions defined as patterns of appraisal variable values Variables evaluated in a fixed sequence Appraisal as a dynamic, evolving process … across multiple modalities … at multiple levels of complexity Conceptual Schematic Perceptual-motor Implementations: Black-box implementations Appraisal dimensions adopted in cognitive-affective architectures

Results of the Appraisal Process: Emotion ‘Specification’ fear .90 probability, importance of affected goals 2 minutes (exp. decay) { aggressive dog | owner} “ aggressive dog approaching” negative { dog | negligent owner | self } low { safety of self | safety of dog | delay } Other appraisal variables….: Type: Descriptive detail: Intensity: Variables affecting intensity: Cause: Direction: Coping potential: Duration: Valence: Goals affected:

Representation & Reasoning Alternatives Vector spaces (Scherer) Connectionist (Velasquez) Symbolic Rules (Marinier, Jones, Henninger, Hudlicka…) Belief nets (de Rosis, Hudlicka, …) Complex symbolic structures (Elliot, Reilly, Gratch & Marsella) Appraisal frames, causal plan structures Spreading activation over networks of processes (Breazeal) Decision-theoretic Decision trees Decision theoretic formulations (Gratch & Marsella, Lisetti & Gmytrasiewicz) Blackboards and ‘specialists’ (Gratch & Marsella) Finite state machines (Kopecek) Markov models (El Nasr) Theorem proving (Zippora) Dynamical systems

Bayesian Belief Networks (MAMID, Hudlicka)

Complex Causal Interpretation (EMA, Gratch & Marsella)

Emotion Effects on Cognition Cognitive-Affective Architecture Stimuli Situations Expectations Goals Affect Appraiser Emotions

Computational Tasks for Modeling Emotion Effects Emotion(s) Cognition Attention, perception, memory, learning, problem-solving, decision-making…) Behavior Verbal, non-verbal, action selection … & other affective factors: Affective States Moods Traits Processes and structures affected Variability in effects (by intensity, by individual…) Integration of multiple emotions (in cognition, in behavior) Effect(s)

Influential Theories Fewer theories exist than for appraisal Specific mechanisms of emotion effects not as well developed Existing theories: Spreading activation & priming (Bower, 1984) … memory effects & biases Distinct modes of processing associated with different emotions (Oatley & Johnson-Laird, 1987) Emotions as patterns of parameters modulating processing (Fellous, Hudlicka, Matthews, Ortony et al., …)

Emotions As Distinct Modes of Processing Parameter-controlled ‘global’ effects across multiple processes Neuromodulation theories (Fellous, 2004) Effects on low-level fundamental processes: attention & working memory Speed & capacity & content bias (e.g., threat, self) Effects on long-term memory Encoding and retrieval: speed & elaboration & bias (threat, self) Effects on higher-level processes Problem-solving, decision-making, planning.. Affect appraiser processes (e.g., assessments of coping potential) Can ‘higher-level’ effects be explained (& implemented ) in terms of effects on the fundamental processes?

Emotions As Parameters (MAMID, Hudlicka) Traits Extraversion Stability Conscientiousness Aggressiveness STATES / TRAITS Processing Structural Module Parameters Construct parameters Architecture topology Long-term memory speed, capacity Cue selection & delay …. Data flow among modules Content & structure Affective States Anxiety Anger Sadness Joy ARCHITECTURE PARAMETERS COGNITIVE ARCHITECTURE Attention Action Selection Situation Assessment Goal Manager Expectation Generator Affect Appraiser

Modeling Threat Bias Processing Parameters Construct parms. - Cue selection - Interpretive biases ... Process Threat cues Process Threatening interpretations Traits Low Stability TRAITS / STATES COGNITIVE ARCHITECTURE PARAMETERS COGNITIVE ARCHITECTURE Attention Action Selection Situation Assessment Goal Manager Expectation Generator Affect Appraiser Emotions Higher Anxiety / Fear Predisposes towards Preferential processing of Threatening stimuli Threat constructs Rated more highly

Modeling Affect-Induced Differences in Behavior MAMID architecture modeled behavior of peacekeeper unit leaders Units encountered a series of ‘surprise’ events en route Hostile crowds Ambushes Destroyed bridges Different leaders defined by distinct personality profiles: “ Normal” leader “ High anxious” leader “ High aggressive” leader Parameter-controlled ‘micro effects’ resulted in observable differences in behavior & distinct ‘mission outcomes’

Distinct Individual Profiles & Behavior “ Normal” “Anxious” Attention Perception / Situation Assessment Expectation Generation Affect Appraisal Goal Selection Action Selection Hostile large crowd Hostile large crowd Objective near Unit capability high Limited # of high-threat & self cues Movement blocked Danger to unit low Danger to unit and self high Perceptual threat & self bias Anxiety: Normal Anxiety: High Rapid-onset of high anxiety Danger from crowd unlikely Danger to unit and self high Career success threatened Threat and self oriented expectations Non-lethal crowd control Reduce anxiety Defend unit Threat and self focus goals Stop Stop; Lethal crowd control Non-lethal crowd control Report info Request help Request info Anxiety regulating behavior

Representation & Reasoning Alternatives Symbolic - specific emotions linked to particular effects & behavior Rules Belief nets Connectionist (Araujo) Parameters bias processing within a network Decision-theoretic (Busemeyer) Decision field theory

What Level of Resolution? “Black box” vs. “Process” models May be all that is required for a particular application Easier to build (…initially) Black box models - simulate input-output mappings ??? INPUT OUTPUT Don’t know and don’t care Stimulus Emotion Emotion Effects

Process Models - emulate internal processing Implement hypothesized mechanisms mediating the I-O mapping Necessary if aiming to understand emotion processes More difficult initially, but more robust and general Cognitive-Affective Architecture INPUT OUTPUT Process #1 Process #2 Process #3 Memory A Memory B Would like to know and do care

What Type of an Architecture? Which architectural components are necessary? Attention, situation assessment, expectation generation, affect appraiser, planner..? Data and control paths among the modules? What fundamental processing paradigm? Sequential see-think-do (see-think/ feel- do?) vs. parallel distributed processing Where does emotion reside within the architecture? Emotion as dedicated modules? vs. emotions as modulating parameters? vs. emotions as emergent properties of a complex, multi-level architecture?

Components of a Cognitive-Affective Architecture : See-Think-Feel-Do “ See” Attention Sensing and Perception “ Think” Situation Assessment Causes Current assessments Future predictions (expectations) Goal management Goals, drives, desires, norms, value Problem solving, Planning, Learning Memory (declarative, procedural, episodic) (sensory, working , long-term) “ Feel” Affect appraiser Emotion effects “ Do” Effectors Performance monitoring

Questions Regarding Representational & Reasoning Requirements What must represented explicitly? Time (present, past, future) Hope needs expectations, regret needs past Mental constructs situations, expectations, goals Memories what type – declarative, episodic, procedural Explicit representation of the self need for complex emotions, social interaction, coping What types of reasoning are necessary? What-if … to generate expectations which influence emotions Causal explanation ..important for attribution

Examples of Cognitive-Affective Architectures Emotion-augmented cognitive architectures Recognition-primed decision-making (Hudlicka) Belief-Desire-Intention architectures (de Rosis…) Soar ( Marinier, Jones, Henninger ) ACT ( Ritter ) Generic - the ‘triune’ architectures Sloman et al., (Cog_Aff) or Sim_Agent (implementation) Leventhal & Scherer (design) Ortony, Normal and Revelle (ONR) (design)

MAMID Cognitive-Affective Architecture Action Selection Cues: State of the world ( “growling dog”, “approaching”) Situations: Perceived state ( “aggressive dog” ) Expectations: Expected state (“dog will attack”, “bite wound”) Goals: Desired state (“protect self”) Actions: to accomplish goals (“climb tree”) Affective state & emotions: Negative valence High anxiety Low happiness Cues Actions Attention Situation Assessment Expectation Generator Affect Appraiser Goal Manager

“ The Triune” Generic Architectures (Sloman; Leventhal & Scherer; Ortony et al.; Arbib & Fellous..) Reactive Routine Reflective hardwired, fixed Well-learned behavior Awareness Compare alternatives - detect deviations Simple ‘what if’ Symbolic processing Approach / Avoid Simple drives Complex mental models Self representations & self-awareness Explicit predictions, causality… Meta-cognition Proto-affect Good/bad Primitive Emotions Good/bad Now/later Full fledged emotions -Basic -Complex flexible

Outline Motivation & Objectives Emotions – Background Info Developing Affective User Models Framework for model analysis Summary & Conclusions

Framework for Development, Analysis and Comparison of Models Which modeling objectives? Which emotions? Which aspects of emotions? (modality, functions, roles) Which processes modeled? (appraisal, effects) Which theory is basis of model? What degree of model resolution Which architecture? (modules, processes, data & control flow) Which representational & reasoning formalisms used? What validation method used?

Summary Need for including (some) emotions in (some) user models Background info from emotion research in psychology and neuroscience Guidelines for development of affective user models & cognitive-affective architectures Requirements for modeling core components of emotions: Cognitive appraisal Emotion effects and emotion-cognition interactions Framework for analysis of computational models of emotion

Successes & State of the Art (1) Research Terminological clarifications Increasing interaction among experimentalists & modelers and theorists Construction of process models of appraisal theories Beginnings of process models of emotion effects Convergence on architecture structure Beginnings of principled analyses of modeling requirements

Successes & State of the Art (2) Research Terminological clarifications Increasing interaction among experimentalists & modelers and theorists Construction of process models of appraisal theories Beginnings of process models of emotion effects Convergence on architecture structure Beginnings of principled analyses of modeling requirements Applications Many ‘shallow’ models enhancing HCI and agents Beginnings of ‘deep’ models driving synthetic agent & robot behavior Emotion sensing & recognition Emotion “expression” Gratch & Marsella De Rosis Breazeal

Challenges Theories to guide model building Appraisal, mechanisms of emotion effects, meta-cognition & emotion Emotion dynamics Multiple emotions & non-linear effects Interaction among multiple modalities Data Emotion experiments are difficult Model data requirements frequently exceed data availability Model development Standards, shared data & ontologies, plug & play modules, guidelines Can we build LTM’s or must they “evolve” through agent-environment interactions (Matthews, 2004) Validation Verification vs. validation Developing validation criteria & benchmark problems

Parting Thought “ Anyone can model emotions. That is easy. But to model emotions - in the right context, - to the right degree, - at the right time, - for the right reason, and - in the right way, this is not easy.” Paraphrasing “On anger”, Aristotle, Nichomachean Ethics

Depth of Feelings: Modeling Emotions in User Models and Agent Architectures

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Depth of Feelings: Modeling Emotions in User Models and Agent Architectures (20)

Recently uploaded (20)

Depth of Feelings: Modeling Emotions in User Models and Agent Architectures