Generating Appraisals with Sequence & Influence Networks Bob Marinier, SoarTech 29 th Soar Workshop June 2009

Motives Appraisal theories define information that is useful to agents How is that information generated? Exploring network representations of sequence and causal knowledge This is mostly speculative; very little has been implemented 2

Background: Appraisal Theories of Emotion Situation Goals Appraisals Emotion Coping 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. Result of appraisals influences emotion Emotion can then be coped with (via internal or external actions) 3

Appraisals to Emotions (Scherer 2001) JoyFearAnger SuddennessHigh/mediumHigh UnpredictabilityHigh Intrinsic pleasantnessLow Goal/need relevanceHigh Cause: agentOther/natureOther Cause: motiveChance/intentionalIntentional Outcome probabilityVery highHighVery high Discrepancy from expectation High ConducivenessVery highLow ControlHigh PowerVery lowHigh 4

Semantic sequence networks Encode knowledge about common sequences  E.g., After breakfast, I always go to work Not episodic – not about specific sequences 5

Example: Thesis sequence network A sequence network consists of nodes corresponding to events, and directed links corresponding to time steps Event 1 Event 2 Event Can create predictions (e.g., If see Event 1, will see Event 3 next) Can generate outcome probability appraisal (e.g., P(Event 3) = 0.4/( ) = 80% Can generate discrepancy from expectation by comparing outcome to prediction Other appraisals computed by separate rules not related to this network Link strength 6

Methods for generating appraisals Thesis workHierarchical sequence networksInfluence networks Suddenness Familiarity Predictability Intrinsic pleasantnessRules Goal relevanceRules Causal agency Causal motive Outcome probabilityNetwork link strength Discrepancy from expectationNetwork node Goal conducivenessRules Urgency Control Power Adjustment Internal standards compatibility External standards compatibility 7

Example: Hierarchical semantic sequence networks Scenes (slots) Hierarchical structure Temporal dependencies Predictions Actions Roles, turn-taking Events 8

Generating appraisal values Knock Event Binding Activate Parent Generate Action Visitor = Guest Greeter = Me Activation = 1.0 Visitor = Solicitor Greeter = Me Activation = 0.5 Generate Expectation Outcome Probability = this interp’s frac of total activation Discrepancy = F(BA, Expectation) Familiarity = Sum(binding activations) Predictability = Change(activation of previously active) Goal Relevance = Change(activation of goal) Goal Conduciveness = +/-Change(activation of goal) Causal Agency = One of bound variables Control = can anyone do action Power = can I do action Causal Motive = property of sequence Urgency = property of temporal links 9

Methods for generating appraisals Thesis workHierarchical sequence networksInfluence networks Suddenness FamiliarityActivation change PredictabilityActivation change Intrinsic pleasantnessRulesEvent/sequence metadata Goal relevanceRulesActivation change Causal agencySequence binding Causal motiveSequence metadata Outcome probabilityNetwork link strengthRelative activation Discrepancy from expectationNetwork nodeBinding activation Goal conducivenessRulesActivation change UrgencySequence metadata ControlRules? Bindings? PowerRules? Bindings? Adjustment Internal standards compatibility External standards compatibility 10

Influence networks Encode knowledge about causal relationships  Nodes are concepts, links are +/- causal relationships Like a specialized form of semantic memory 11

Example: Influence networks Impress Boss Be Healthy Quit Smoking + Importance = Be Successful Importance = 1 Exercise + Cold Turkey Patch Cold Turkey ……… + + Stress + Be Happy - Importance = 1 Possible Connection to Sequences Goals and Values Concepts Causal Relationships 12

Generating appraisal values Impress Boss Be Healthy Quit Smoking + Importance = Be Successful Importance = 1 Exercise + Cold Turkey Patch Cold Turkey ……… + + Stress + Be Happy - Importance = 1 Likelihood ↑ Likelihood ↓ Goal Conduciveness = Weighted avg of likelihood from considered “action” at goals Outcome Probability = Inverted Likelihood Control and Power = Ability to implement Likelihood ↑ Inv Likelihood ↑ Adjustment = Weighted avg of likelihood from implementation at goals Likelihood ↑ Likelihood ↓ Standards Compatibility = Weighted avg of likelihood at values 13

Methods for generating appraisal values 14 Thesis workHierarchical sequence networksInfluence networks Suddenness FamiliarityActivation change PredictabilityActivation change Intrinsic pleasantnessRulesEvent/sequence metadata Goal relevanceRulesActivation change Causal agencySequence binding Causal motiveSequence metadata Outcome probabilityNetwork link strengthRelative activationInverted likelihood from goal to considered “actions” Discrepancy from expectationNetwork nodeBinding activation Goal conducivenessRulesActivation changeWeighted avg likelihood from considered “action” at goals UrgencySequence metadata ControlRules? Bindings?Bindings to sequences? PowerRules? Bindings?Bindings to sequences? AdjustmentWeighted avg likelihood from considered “implementation” at goals Internal standards compatibilityWeighted avg likelihood at values External standards compatibilityWeighted avg likelihood at values

Nuggets and Coal Have a story about where all of Scherer’s appraisals might come from Have multiple stories for some appraisals Actually implemented conduciveness in causal networks Story is mostly hand-wavy Have multiple stories for some appraisals Involves integrating two kinds of networks 15 Suggests new or modified architectural mechanisms to support these kinds of networks?