Computational Models of Emotion and cognition

1. Computational Models of Emotion and cognition Christopher L. Dancy, Frank E. Ritter, Keith Berry Jerry Lin, Marc Spraragen, Michael Zyda Advances in Cognitive Systems Presenter: Yoon-hyungChoi 2014. 09. 11

2. Contents Introduction and Background Representative Models and their Properties Example Models Open Issues Challenges to Near-Term Research Conclusion

3. 01 Introduction and Background • Emotion and Cognition have long been thought to have important interaction - However, there are still open questions and difficulties • There is much confusion regarding emotion terminology. - Affect - Appraisal - Cognition - Emotion - Feeling - Mood

4. 01 Introduction and Background • Computational models of emotion and cognition are those that try to explain emotion - in the context of its intimate relationship with cognition - distinguished from those in psychology and cognitive science • In the 19th century, William James and others theory. - focus on physiological reactions - Appraisal theory(emotion as effects of reactions to situations)

5. 01 Introduction and Background • Appraisal Theory - developed as a means to predict individual human emotions given particular situations - A person can appraise an event, situation with respect to appraiser’s beliefs, desires, and intentions - Most appraisal theories share fundamental concepts : Valence and Arousal • The concept of coping potential - ability to deal with a situation either by action or cognition - “primary” vs. “secondary” (Lazarus, 1966) - “action tendencies” (Frijda, 1987) • The Ortony, Clore, and Collins(1988), OCC theory - categorizes emotion based on appraisal of pleasure / displeasure(valence) and intensity(arousal)

6. 01 Introduction and Background • Non-cognitive vs. Cognitive appraisal - There is not a clear line • PAD dimensional theory(Marsella & Gratch, 2009) - dimensions of Pleasure, Arousal, and Dominance - analogous to coping potential in appraisal theory - “anger” vs. “anxiety”

7. 02 Representative Models and their Properties • The impact of recent emotion-related human psychological and cognitive studies - how thinking of a plan changes affective state - focus on basic expressions to register the presence of emotions - “behavior-consequent”, “cognitive-consequent” • Well-defined base cognitive theory or integrated cognitive model - EMA, Soar-Emote, ACT-R • Effects are often cast as constraints on goal and action choices(i.e., decisions) - EMA - Meyer’s system • Decision biases - WASABI(Becker-Asano’s, 2009) - BehBehBeh and other models of Frijda’s theory such as ACRES/WILL(Moffat, Frijda, & Phaf, 1993)

8. 02 Representative Models and their Properties • Emotional biases on learning - typically memory based - reinforce recall and decision biases • Emotion as a recall heuristic has been handled in different ways - ACT-R with its well-tested model of associative memory - MAMID models emotional effects on cognitive recall and inference

9. 02 Representative Models and their Properties

10. 03 Example Models • EMA - Appraisal frame - Emotional process and timing issues with cognition - Effects on the primary cognitive activities of planning and inference have not been demonstrated - only a few simple configurations have been demonstrated • Soar-Emote(PEACTIDM) - an explicit model of cognitive control and Scherer’s appraisal theory - only one appraisal frame of significance in the system per cognitive cycle - how appraisals may be integrated in the cognitive cycle - how to calculate both arousal and valence in various model - translating an appraisal frame into a discrete emotion does not make sense

11. 03 Example Models • WASABI - one of the most general models of emotion - modeling of primary emotions and secondary emotions - grounding of secondary emotions in primary in-born emotions - lossy compression of information

12. 04 Open Issues • Criteria and methods for model evaluation - emotion likely has an intimate relationship with nearly all components of cognitive architecture - compare the model in its ability to perform like a human(Gratch & Marsella, 2005) - Encode a corpus of emotional situations within a model and compare results (Gratch, Marsella, &Mao, 2006) • The issue is that of the domain’s complexity and breadth - we cannot shy away from complexity - researchers must deal with many topics - encourage the creation of a collaborative community based around an open flexible architecture.

13. 05 Challenges to Near-Term Research • There are several important challenges - moving towards uniformity in emotional representations and mechanisms - understanding existing use of emotions in traditional artificial intelligence - exploring innovative uses of emotions and emotion engineering • There is a clear deficit in models - they suffer from either being narrowly focused on one mechanism of interaction - explicit study and modeling should be fundamental • There is opportunity to unlock answers to difficult problems in AI

14. 06 Conclusion • We analyzed their levels of emotional-cognitive integration - to help understand how each system compares and contrasts with others - identified several key properties of models • We identified the significant open issues - standardizing criteria for evaluation of models - The complexity and breadth of the domain - integration with the rich history of AI research

15. Thank you!