1. Emotion-Driven Reinforcement Learning Bob Marinier & John Laird University of Michigan, Computer Science and Engineering CogSci’08

2. Introduction Interested in the functional benefits of emotion for a cognitive agent Appraisal theories of emotion PEACTIDM theory of cognitive control Use emotion as a reward signal to a reinforcement learning agent Demonstrates a functional benefit of emotion Provides a theory of the origin of intrinsic reward 2

3. Outline Background Integration of emotion and cognition Integration of emotion and reinforcement learning Implementation in Soar Learning task Results 3

4. Appraisal Theories of Emotion A situation is evaluated along a number of appraisal dimensions, many of which relate the situation to current goals Novelty, goal relevance, goal conduciveness, expectedness, causal agency, etc. Appraisals influence emotion Emotion can then be coped with (via internal or external actions) Situation Goals Appraisals Coping Emotion 4

5. Appraisals to Emotions (Scherer 2001) 5

6. Cognitive Control: PEACTIDM (Newell 1990) 6

7. Unification of PEACTIDM and Appraisal Theories 7 Perceive Raw Perceptual Information Environmental Change Encode Motor Suddenness Unpredictability Goal Relevance Intrinsic Pleasantness Stimulus Relevance Motor Commands Prediction Outcome Probability Attend Decode Causal Agent/Motive Discrepancy Conduciveness Control/Power Stimulus chosen for processing Action Comprehend Intend Current Situation Assessment

8. Distinction between emotion, mood, and feeling(Marinier & Laird 2007) Emotion: Result of appraisals Is about the current situation Mood: “Average” over recent emotions Provides historical context Feeling: Emotion “+” Mood What agent actually perceives 8

9. Emotion, mood, and feeling Cognition Active Appraisals Perceived Feeling Emotion Feeling Combination Function Pull Mood Decay 9

10. Intrinsically Motivated Reinforcement Learning(Sutton & Barto 1998; Singh et al. 2004) 10 External Environment Environment Actions Sensations Critic “Organism” Internal Environment Actions States Rewards Critic Appraisal Process Agent +/- Feeling Intensity States Rewards Decisions Agent Reward = Intensity * Valence

11. Extending Soar with Emotion(Marinier & Laird 2007) Episodic Semantic Symbolic Long-Term Memories Procedural Semantic Learning Episodic Learning Chunking Reinforcement Learning Appraisal Detector Short-Term Memory Situation, Goals Decision Procedure Visual Imagery Perception Action Body 11

12. Extending Soar with Emotion(Marinier & Laird 2007) 12 Episodic Semantic Symbolic Long-Term Memories Procedural Semantic Learning Episodic Learning Chunking Reinforcement Learning +/-Intensity Appraisal Detector Feeling .9,.6,.5,-.1,.8,… Short-Term Memory Situation, Goals Feelings Decision Procedure Feelings Appraisals Visual Imagery Emotion .5,.7,0,-.4,.3,… Mood .7,-.2,.8,.3,.6,… Perception Action Knowledge Body Architecture

13. Learning task Start Goal 13

14. Learning task: Encoding 14 North Passable: false On path: false Progress: true East Passable: false On path: true Progress: true West Passable: false On path: false Progress: true South Passable: true On path: true Progress: true

15. Learning task: Encoding & Appraisal 15 North Intrinsic Pleasantness: Low Goal Relevance: Low Unpredictability: High East Intrinsic Pleasantness: Low Goal Relevance: High Unpredictability: High West Intrinsic Pleasantness: Low Goal Relevance: Low Unpredictability: High South Intrinsic Pleasantness: Neutral Goal Relevance: High Unpredictability: Low

16. Learning task: Attending, Comprehending & Appraisal 16 South Intrinsic Pleasantness: Neutral Goal Relevance: High Unpredictability: Low Conduciveness: High Control: High …

17. Learning task: Tasking 17

18. Learning task: Tasking 18 Optimal Subtasks

19. What is being learned? When to Attend vs Task If Attending, what to Attend to If Tasking, which subtask to create When to Intend vs. Ignore 19

20. Learning Results 20

21. Results: With and without mood 21

22. Discussion Agent learns both internal (tasking) and external (movement) actions Emotion allows for more frequent rewards, and thus learns faster than standard RL Mood “fills in the gaps” allowing for even faster learning and less variability 22

23. Conclusion & Future Work Demonstrated computational model that integrates emotion and cognitive control Confirmed emotion can drive reinforcement learning We have already successfully demonstrated similar learning in a more complex domain Would like to explore multi-agent scenarios 23

24. 24 HIGH INTENSITY alert tense excited nervous elated stressed happy upset NEGATIVE VALENCE POSITIVE VALENCE sad contented depressed serene lethargic relaxed fatigued calm LOW INTENSITY Circumplex models Emotions can be described in terms of intensity and valence, as in a circumplex model: Adapted from Feldman Barrett & Russell (1998)

25. Computing Feeling from Emotion and Mood 25 Assumption: Appraisal dimensions are independent Limited Range: Inputs and outputs are in [0,1] or [-1,1] Distinguishability: Very different inputs should lead to very different outputs Non-linear: Linearity would violate limited range and distinguishability

26. Computing Feeling Intensity 26 Motivation: Intensity gives a summary of how important (i.e., how good or bad) the situation is Limited range: Should map onto [0,1] No dominant appraisal: No single value should drown out all the others Can’t just multiply values, because if any are 0, then intensity is 0 Realization principle: Expected events should be less intense than unexpected events