1. A Theory of Unifying Cognitive Processing, Appraisal and Emotion Bob Marinier John Laird University of Michigan

2. 2 Overview Introduction Emotions Agent Processing Unification

3. 3 Introduction Have independent theories of agent processing and emotion Agent Processing: Allen Newell’s PEACTIDM Emotion: Appraisal Theory Each of these is incomplete Emotions and rationality are tightly integrated in humans Emotions prepare humans for situations and motivate them, but still allow for flexibility in generating responses (i.e. can do more than react) How can we unify agent processing and emotion?

4. Emotion: Appraisal Theory

5. 5 Appraisal Theory of Emotion Suppose a person has some goals, beliefs, etc. (knowledge) An event occurs (internal or external) The person appraises the relationship between his goals and the event along a number of dimensions (e.g. unexpectedness, conduciveness, agency, etc). The appraisal automatically leads to emotion (e.g. physiological/cognitive changes, thought-action urges, etc) The person perceives emotion as feelings (internal event) The person copes with feelings by taking internal or external actions to improve/maintain the relationship between his goals and the environment

6. 6 Scherer 2001Roseman 2001 Smith & Lazarus 1990; Smith & Kirby 2001 Lazarus 1991/2001 Gratch & Marsella (2004) Novelty: Suddenness Novelty: Familiarity Novelty: Predictability Intrinsic pleasantness Goal/need relevanceMotivational relevanceGoal relevanceRelevance Cause: agent AgencySelf/Other accountabilityBlame and credit Causal attribution Cause: motive Outcome probability Probability Future expectancyFuture expectationsLikelihood Urgency Discrepancy from expectation Unexpectedness ConducivenessSituational stateMotivational congruenceGoal congruenceDesirability Control Control potential Problem-focused coping potential Coping potential Changeability PowerControllability Adjustment Emotion-focused coping potential Internal standards compatibility Type of ego involvement Perspective External standards compatibility Motivational state Problem type 14/Complete17/Complete5/Partial17/Partial6/Partial

7. 7 Scherer 2001Elation/JoyFearRage/Hot Anger SuddennessHigh/mediumHigh FamiliarityLow PredictabilityLow Intrinsic pleasantnessLow Goal/need RelevanceHigh Cause: agentOther/natureOther Cause: motiveChance/intentionalIntentional Outcome probabilityVery highHighVery high Discrepancy from ExpectationDissonant ConducivenessVery highObstruct UrgencyLowVery highHigh ControlHigh PowerVery lowHigh AdjustmentMediumLowHigh Internal standards compatibility External standards compatibility Low Appraisals to Emotions Relevance Implication Coping potential Normative Significance

8. 8

9. 9 What’s missing? How are appraisals generated? When are appraisals generated? How do emotion and appraisal impact behavior (in detail)?

10. Agent Processing: PEACTIDM And unification with appraisal and emotion

11. 11 Agent Processing: Allen Newell’s PEACTIDM PerceiveRaw perception EncodeDomain-independent representation AttendChose stimulus to process ComprehendGenerate a structure that relates stimulus to goals that can be used to inform behavior TaskingPerform goal maintenance IntendChose an action DecodeDecompose action into motor commands MotorExecute motor commands Event Processing Response Processing

12. 12 Event Processing: Desirable Properties Domain independent Limited working memory Happens over time Incremental Supports immediate comprehension Supports hierarchical comprehension Supports prediction Influenced by external processes

13. 13 Encode will generate domain-independent structures from the raw information it gets from Perceive Possible encoded structures Talmy (1975) Figure Motion Path Ground Simplification Actor Action May also include information about novelty (e.g. this is a common event or not, it occurred suddenly or not, etc) Encode and Event Structure Bob Walking Across Street Bob Walking across street

14. 14 Attend There may be multiple events occurring simultaneously Encode is a fast parallel process, and thus all perceived events may be encoded Attend picks one to focus on and comprehend next May pick based on novelty, which is generated by Perceive and Encode

15. 15 Comprehension Process Goal: To create data structures that inform behavior Key: Process sequences of events Process Observe some sequence of events (e.g. the beginning of some sequence) Match partial sequence to known complete sequence Use complete sequence to predict next event Only work on one event or sequence at a time (i.e. processing is local) Since the event structures are domain independent, this process is also domain independent

16. 16 Abstract Events, Sequences and Subgoals An event sequence can be abstracted to represent a single event in a more abstract sequence Example: Step down from curb Take a few steps Step up onto curb …this is just the “Cross the Street” event, which may be just one event in the “Get from Car to Office” sequence, which may be one event in the “Go to Work” sequence…which may be just one event in the “Living My Life” sequence. Abstract events can be thought of as subgoals There are many possible ways to represent these hierarchies….

17. 17 Step Down Walk Cross the Street Walk up to Building Open Door Get From Car to Office Event Knowledge Hierarchy 1 Enter Building Cross the Street Enter Building Get From Car to Office Events Sequences Events Sequences Events Step Up Go to Work … Sequences Events

18. 18 Go to Work 1 Go to Work 2 … AB 1 AB 2 … Step Down Walk Enter Street Step Up Walk up to Building Cross the Street CS & AB ES 1 ES 2 … Cross the Street 1 Cross the Street 2 … CS & AB 1 CS & AB 2 … Events Sequences Events Sequences Events Event Knowledge Hierarchy 2 Leave Street LS 1 LS 2 …EB1 1 EB1 2 … EB1 Approach Building EB2 1 EB2 2 … Open Door EB2 Enter Building 1 Enter Building 2 … Get from Car to Office 1 Get from Car to Office 2 … AB & EB 1 AB & EB 2 … Enter Building AB & EB Get from Car to Office Sequences Events Go to Work Sequences Events ……

19. 19 Comprehension Process Details Comprehend Event Reinterpret Events Matches Prediction? Determine Probability Determine Causality Determine Goal Conduciveness Extract Information to Inform Behavior YES NO Appraisals

20. 20 Event Processing: Desirable Properties Revisited Domain independentEvents are domain-independent Limited working memoryOne interpretation at a time Happens over timeEvents occur over time IncrementalAttend to one event at a time; Local processing Supports immediate comprehension Can always guess at complete sequence based on event Supports hierarchical comprehension Sequences can be abstracted to events Supports predictionNext event can be read from guessed complete sequence Influenced by external processes Ambiguity resolution can be biased by current goal, emotion, memory activation, etc.

21. 21 Agent Processing: Allen Newell’s PEACTIDM PerceiveRaw perception EncodeDomain-independent representation AttendChose stimulus to process ComprehendGenerate a structure that relates stimulus to goals that can be used to inform behavior TaskingPerform goal maintenance IntendChose an action DecodeDecompose action into motor commands MotorExecute motor commands Event Processing Response Processing

22. 22 Tasking Process Goal: Update current goal if threatened Key: Emotion automatically signals with status (goal threatened, situation alterable) and how to fix it (e.g. whose fault is it, etc) Process: If feel bad (i.e. goal threatened) If feel angry (e.g. can alter course of events), set new subgoal to get back on track If feel afraid (e.g. can’t alter course of events), give up on goal

23. 23 Tasking Process Details Tasking Set New Subgoal Feel Bad? Set New Goal Angry How? YES Afraid Feelings Coping Strategies

24. 24 Intend Process Goal: Determine next action to execute Key: In general, there may be many paths from the current situation to the goal, so Intend must pick one In general, may also have to contend with action tendencies (e.g. automatic responses) Process: If urgency is high, chose “automatic” response Otherwise, walk event hierarchy to find path to goal If first event on path is mine, execute it (otherwise wait for event to occur)

25. 25 Scherer 2001Generated ByRequired By Novelty: SuddennessPerception Attend Novelty: Familiarity Encoding Novelty: Predictability Intrinsic pleasantness Tasking (via Feelings) Goal/need relevance Comprehension Cause: agent Cause: motive Outcome probability Urgency Discrepancy from expectationComprehension Conduciveness Tasking (via Feelings) Control Power Adjustment Internal standards compatibility External standards compatibility Unification

26. 26

27. 27 Predictions Agent should be interruptible Partial ordering constraint on appraisal generation Different emotions may require different amounts of processing Time constraints may lead to errors in comprehension (and thus emotion)

28. 28 Future Work Tracking multiple goals Keeping track of others’ goals Better-than-random disambiguation Dealing with loosely-connected event hierarchies

29. 29 Summary: Unification of PEACTIDM, Appraisal and Emotion PEACTIDM defines critical functions, but not how they are achieved, e.g. What is generated by comprehension? How does tasking work? Appraisal and Emotion define critical data used to drive behavior (and which of those data arise from automatic or knowledge- driven processes), but not how the data are generated by the agent’s processing, e.g. When is appraisal generated? Why is appraisal generated then? In general, Need appraisal information in order to Attend, Comprehend, Task and Intend Comprehension (and Perception and Encoding) provide this appraisal information Emotion summarizes it (and also generates common automatic responses and prepares body to take those actions)