1. (How) Can Appraisal Theory be Formalized at a Meta-level? Joost Broekens, Doug DeGroot LIACS, Leiden University

2. Why formalize appraisal structure at high level? Appraisal theory development. –Comparison, refinement, convergence Architectural basis for computational models –Development and debugging.

3. Emotions in Agents What is an emotion? –Heuristic relating events to goals, needs, desires, beliefs of an agent (cognitive definition). –Communication medium. –Related to homeostasis and hormonal state Why use an emotion in agents and robots? –heuristic aspect (efficient evaluation), communicative aspect. Which agents might need emotions? –Games, HCI, HRI, Virtual-Reality, Decision-making and planning. Computational models of emotion, in general, are based on Cognitive Appraisal Theory.

4. Structural Theories (what), Process Theories (how) Structural Theory: structural relation between: –Environment of agent (perception) –Appraisal processes that interpret the environment in terms of values on appraisal dimensions (appraisal) –Mediating processes that relate appraisal dimension values to emotions (mediation) –Processes are black-boxes. –Declarative semantics Process Theory: –Detailed cognitive operations and mechanisms involved in processes and their interaction as described by structural theory of appraisal. –Procedural (cognitive) semantics

5. Computational models of Emotions Structural Theory + assumptions from AI = computational model (Gratch and Marsella, 2004). This poses a problem (Gratch and Marsella, 2004) –Structural Appraisal Theory: abstract. –Computational model: algorithmic, detailed. What if the model does something unexpected? Structural Theory Computational Model AI Assumptions Gap

6. What’s wrong? The Computational Model or the Theory (or the observer)? Structural Theory Computational Model Situation AI Assumptions ? Gap Formal Description Computational Model Situation Assume less :-) Structural Theory Smaller Gap

7. Problem: How to Debug Your Computational Model? Debugging is a problem: –Large gap between theory and computational model. –Highly complex agent designs complicate debugging. –Understanding emotions is not something computer scientist are trained, in contrast it’s the appraisal theorist’s job.

8. Benefits of Such Formalisms Appraisal Theory –Comparison, Integration, Convergence (Wherle and Scherer, 2001) –Precise and structured theory revision –Process of Formalization helps theory development and refinement. –Formal annotation of experimental results. Computational models –Formal architecture of appraisal. –Evaluation of computational model in relation to the theory –Structured storage of annotated experimental results (human/agent) Compare computational models. Feedback to theory and human-subject based experimental results

9. Requirements for a Formalism for the Structure of Appraisal How many, which processes exist (perception, appraisal, mediation) When and how are these activated (threshold, continuous?) How much time needed to evaluate? What kind of information needed for these processes? How many and which appraisal dimensions, emotional response components? How do appraisal dimension values relate to emotional response components? See also (Reisenzein, 2001).

10. Overview of the Formalism (1/4): Perception W = observable objects and events in the environment of the agent P = the set of all perception processes available to the agent. p i :W n  V n  I n  O n i. Is a perception process translating the world into mental objects ( O ) in the context of a current emotion ( I ) and appraisal state ( V ). O = set of all mental objects currently perceived by the agent with PAM External World (W) Appraisal Dimension Values (V) Emotion Component Intensities(I) Mental Objects (O)

11. Overview of the Formalism (2/4): Appraisal A = the set of appraisal processes. a i :O n  I n  V i n, a i is an appraisal process, mapping mental objects ( O ) to possible appraisal dimension values ( V ) in the context of the current emotion ( I ). D = set of appraisal dimensions defined by the theory. V = set of current appraisal dimension values V  O n  D  [-1,1] PAM External World (W) Appraisal Dimension Values (V) Emotion Component Intensities(I) Mental Objects (O)

12. Overview of the Formalism (3/4): Mediation E = set of possible emotional response components I = set of emotional response component intensities I=I  E  [0,1] M = set of mediating processes. m j :V n  I j is a mediating process relating appraisal dimension values ( V ) to emotional component intensities ( I ) PAM External World (W) Appraisal Dimension Values (V) Emotion Component Intensities(I) Mental Objects (O)

13. Overview of the Formalism (4/4): Process dependencies PP = set of all processes ( P, A and M ) LT = set of process dependency types. G = set of guards L = set of process dependencies. L = PPxPPxGxLT (  x)(  y) processing in q x is influenced iff ((p y,q x,g,n)  L  g=true  p,q  PP  g  G  n  N) If a dependency exists between a process p and q and the guard g of that link is true, processing in q is influenced in a way denoted by the type n PAM External World (W) Appraisal Dimension Values (V) Emotion Component Intensities(I) Mental Objects (O)

14. Formalization of structure Appraisal theory development. Architectural basis for computational models

15. Application 1: Integration of two Appraisal Theories. Integration based on: –Scherer’s Stimulus Evaluation Checks (SEC) (Scherer, 2001) –Smith and Kirby’s Appraisal Detector Model (ADM) (Smith and Kirby, 2000) SEC: multiple appraisal processes (stimulus checks) –Appraisal Processes activate in four* consecutive steps: Relevance detection, Implication assessment, Coping potential, Norm/self compatibility. –Processes exist at three perception levels: sensory-motor, schematic, conceptual. –Current result of appraisal processes stored in appraisal registers. * Here we only use the first three.

16. Application 1: Integration ADM: –Appraisal detectors integrate appraisal information coming from different perception levels (levels equivalent to those defined in SEC, i.e., sensory-motor, schematic, conceptual) –Appraisal detectors produce emotional response. –Feedback from emotional response to processing, specifically conceptual (reasoning) and schematic (associative learning) levels. Integration basics: common architectural concepts –Separation of appraisal in three levels of information processing. –Appraisal registers/detectors

17. Application 1:Integration SuddennessFamiliarityPredictability Stimulus perception Intrinsic pleasantness Schematic Conceptual Goal/Need relevance Causal attribution Agency Intentional attribution Outcome probability Adjustment Expectation discrepancy Goal/Need conduciveness Urgency PowerControl Implication detector Relevance detector Novelty Coping pot. detector

18. Application 1:Integration SuddennessFamiliarityPredictability Stimulus perception Intrinsic pleasantness Schematic Conceptual Goal/Need relevance Causal attribution Agency Intentional attribution Outcome probability Adjustment Expectation discrepancy Goal/Need conduciveness Urgency PowerControl Implication detector Relevance detector Novelty Coping pot. detector

19. Application 1:Integration SuddennessFamiliarityPredictability Stimulus perception Intrinsic pleasantness Schematic Conceptual Goal/Need relevance Causal attribution Agency Intentional attribution Outcome probability Adjustment Expectation discrepancy Goal/Need conduciveness Urgency PowerControl Implication detector Relevance detector Novelty Coping pot. detector

20. Application 1:Integration SuddennessFamiliarityPredictability Stimulus perception Intrinsic pleasantness Schematic Conceptual Goal/Need relevance Causal attribution Agency Intentional attribution Outcome probability Adjustment Expectation discrepancy Goal/Need conduciveness Urgency PowerControl Implication detector Relevance detector Novelty Coping pot. detector

21. Application 1:Integration SuddennessFamiliarityPredictability Stimulus perception Intrinsic pleasantness Schematic Conceptual Goal/Need relevance Causal attribution Agency Intentional attribution Outcome probability Adjustment Expectation discrepancy Goal/Need conduciveness Urgency PowerControl Implication detector Relevance detector Novelty Coping pot. detector

22. Application 1:Integration SuddennessFamiliarityPredictability Stimulus perception Intrinsic pleasantness Schematic Conceptual Goal/Need relevance Causal attribution Agency Intentional attribution Outcome probability Adjustment Expectation discrepancy Goal/Need conduciveness Urgency PowerControl Implication detector Relevance detector Novelty Coping pot. detector

23. Application 2: Formal Description of a Computational Model Formal description: –Based on simplified version of integrated model (SSK) –Used to define the architecture of appraisal (i.e., appraisal steps, appraisal detectors, levels of perception, appraisal dimensions) –Used to evaluate behavior of resulting computational model of emotions. Test environment: PacMan –Appraisal of events in PacMan’s environment is simulated. –Architecture and appraisal dimensions used based on simplified SSK model

24. Application 2: Comp. Model

25. Formal description helped to verify model’s behavior. No activation of relevance detection… –Due to bipolar variable: conductiveness. –Summing negatively conductive and positively conductive events results in no conductivity activation  not plausible. Separate conductiveness in pos and neg. –Relevance detection active and activation of implication checks at right moments.

26. Some Conclusions Formal description facilitated development of computational model. –Clear definition of architecture of appraisal processes Formalism facilitated integration of theories. Open: –How to formally encode experiments and experimental results, comparing experimental results, etc. –What is the relation between BDI-based formalism and Meta-level formalisms.

27. Questions? Perception appraisal and mediating processes: P={stimulus_perception, schematic} A={suddenness, familiarity, novelty, intrinsic_pleasantness, relevance, conductiveness, urgency, control, power} M={relevance_detector, implication_detector, coping_potential_detector} Mental object types, mental objects, appraisal dimensions and emotion components: OT={belief} PO={(see_ghost, belief), (lost_ghost, belief), (eaten_by_ghost, belief), (see_edible_ghost, belief), (lost_edible_ghost, belief), (eaten_ghost, belief), (see_power, belief), (eaten_power, belief), (see_dot, belief), (eaten_dot, belief), (see_fruit, belief), (lost_fruit, belief), (eaten_fruit, belief)} D={novelty_dim, intrinsic_pleasantness_dim, conductiveness_dim, urgency_dim, control_dim, power_dim} E={} Link types, guards, data constraints and dependencies: LT={activation} G={true, guard1, guard2, guard3} with: guard1=((  x,y,z) x,y,z  V  x=(o,d,i)  y=(o,d',j)  z=(o,d'',k)  (i+j+k)/3>.15  d=novelty_dim  d'=intrinsic_dim  d''=relevance_dim) guard2=((  x,y) x,y  V  x=(o,d,i)  y=(o,d',j)  (i+j)/2>.25  d=conductiveness_dim  d'=urgency_dim) guard3=((  x,y) x,y  V  x=(d,i)  y=(d',j)  i*j>0  d=control_dim  d'=power_dim) H={c1, c2} with: c1= ((  x)x  V  x=(y,d,i,t)  i>0) if ((  y)y  O  y=(c,j,t')  j>0  t=t'), and c2= ((  z)z  I  z=(e,i',t'')  i'>0) if ((  x')x'  V x'=(y',d,j',t''')  j'>0  t''=t''')}. L={(stimulus_perception, suddenness, true, activation), (stimulus_perception, intrinsic_pleasantness, true, activation), (stimulus_perception, relevance, true, activation), (stimulus_perception, conductiveness, true, activation), (stimulus_perception, urgency, true, activation), (stimulus_perception, power, true, activation), (schematic, familiarity, true, activation), (schematic, relevance, true, activation), (schematic, conductiveness, true, activation), (schematic, urgency, true, activation), (schematic, control, true, activation), (suddenness, novelty, true, activation), (familiarity, novelty, true, activation), (novelty, relevance__detector, true, activation), (intrinsic_pleasantness, relevance_detector, true, activation), (relevance, relevance_detector, true, activation), (relevance_detector, conductiveness, guard1, activation), (relevance_detector, urgency, guard1, activation), (conductiveness, implication_detector, true, activation), (urgency, implication_detector, true, activation), (implication_detector, control, guard2, activation), (implication_detector, power, guard2, activation), (control, coping_potential_detector, guard3, activation), (power, coping_potential_detector, guard3, activation)} Referred literature: Reisenzein, Rainer. Appraisal Processes Conceptualized from a Schema-Theoretic Perspective: Contributions to a Process Analysis of Emotions. 2001. Smith, Craig A., Kirby, Leslie D. Consequences require antecedents: Toward a process model of emotion elicitation. 2000. Wherle, Thomas and Scherer, Klaus R. Towards Computational Modeling of Appraisal Theories. 2001. Scherer, Klaus R. Appraisal Considered as a Process of Multilevel Sequential Checking. 2001. Gratch, Jonathan and Marsella, Stacy. A domain independent framework for modeling emotion. 2004. Broekens and DeGroot,. Formal Models of Emotion: Theory, Specification and Computational Model. 2004