1March 2004 © Kathleen M. Carley – CASOS – ISRI - CMU ActionsKathleen M. CarleyMarch 2004 © Kathleen M. Carley – CASOS – ISRI - CMUProblemPredict future directions for complex events  When is a strike likely to occur When will a riot turn violent What type of attack is al-Qa’ida likely to try next Who will be assassinated nextExplain past events Why did X happen2March 2004 © Kathleen M. Carley – CASOS – ISRI - CMUMake this harder In particular predict and explain events where individual choice and social action are not filtered through the lens of votes or money Votes and money are “easy” Votes – preferences and combining rules Money – optimization framework and tradeoffs In both cases – analysis is facilitated byHaving a uniform scale for evaluationHaving numeric scales – direct comparison is possibleFeedback is relatively quickEvaluation criteria can be objective Consequently it is relatively straight forward to construct aggregation rulesMarch 2004 © Kathleen M. Carley – CASOS – ISRI - CMUExemplar workSee work of TsandholmUsing computational economics3March 2004 © Kathleen M. Carley – CASOS – ISRI - CMUAnd then there are physical actions …Physical movement entails other problemsRecognition of environmentAwareness of locationPhysical limitations…This is hard – but, to an extent, well definedMarch 2004 © Kathleen M. Carley – CASOS – ISRI - CMUExemplar workSee work by VelosoSee robo soccerSee Blocks-world and work by Pat Winston4March 2004 © Kathleen M. Carley – CASOS – ISRI - CMUMany types of actions are harder still (or less understood)Consider actions like support takeover bid laughOften no scale, let alone uniform scale for evaluation Numeric scales for comparing strength of actions at different times and by different actors is difficult scales may differ by individuals so not clear how to “combine” actions across individuals to get group effects – what are principles of aggregationThe impact of actions may  be invisible occur at different time scalesEvaluation criteria are often subjectiveMarch 2004 © Kathleen M. Carley – CASOS – ISRI - CMUApproach 1: Story UnderstandingIdea - there are canonical actions Self-loco-motion, mechanical-loco-motion, speech act, ...Create a complete list of all such actionsDefine a template (schema, frame, …) for each action Self-loco-motionActor – valueActed upon – noneGoal – valueSpeed – valueDirection – valueType – valueDifferent actions can be characterized by filling different valuesExemplar work – Allan Collins5March 2004 © Kathleen M. Carley – CASOS – ISRI - CMUAction SchemasDefine characteristics of action – what happensIdentifies actor (acting)Identifies object (acted upon)Places qualifiers on specific instantiationMay place constraints on actionMay set action pre-requisitesMay set temporal constraintsMarch 2004 © Kathleen M. Carley – CASOS – ISRI - CMUActs as SchemasWorks well in small well defined domainsWorks when an ontology can be well defined and is shared by system usersThis gets complicated when: the impact of an action varies with who the acting and acted upon arethe same action can meet several goalsactions require multiple actors to coordinate actionactions are perpetrated on multiple others and the impact on each can be differentoutcomes are probabilisticThere are approaches – just it’s an unsolved problem6March 2004 © Kathleen M. Carley – CASOS – ISRI - CMUUsing schema’s to predict actionLocating behavioral or historical regularities for schemas is dauntingBuilding cross-cultural schemas may not be possibleDoesn’t explain individual differences in what actions are takenDoesn’t explain outcome differences due to the same actionMarch 2004 © Kathleen M. Carley – CASOS – ISRI - CMUApproach 2: Expert systemIdea – there are condition-action pairings that govern behaviorCreate a list of rules that link conditions to actionsComplications These rules may or may not be probabilistic May or may not be complex (multiple sub conditions may co-exist) May or may not require temporal reasoning (but that’s hard)Choices determine the type of inference engine usedFurther two approaches All actions are possible and rules eliminate possibilities Generate only those actions that match possibilities7March 2004 © Kathleen M. Carley – CASOS – ISRI - CMURules and ActionsWorks well in complex knowledge intensive domainThis gets complicated when The same conditions imply multiple actions Rules become contradictory Knowledge engineering techniques Can’t find rulesRules are actor dependentRules are object dependentRules require an emotional pre-requisiteDifferent scenarios generate seemingly contradictory rulesMarch 2004 © Kathleen M. Carley – CASOS – ISRI - CMUUsing rules to predict actionLocating behavioral or historical regularities for schemas is dauntingBuilding cross-cultural rules may not be possibleIndividual differences in what actions are taken can be handled by making actors objects in the expert systems statespace – Requires class of actors or look up tables for individualsDoesn’t explain outcome differences due to the same action8March 2004 © Kathleen M. Carley – CASOS – ISRI - CMUApproach 3: Goal based reasoningIdea – individuals take actions to meet goalsOptimization procedures are used to locate action most likely to achieve goalThis gets complicated when Actions are not readily placed in a numeric framework  Goals are not easily articulated Actions cannot be directly related to goals Emotions, location, etc. alter the relation of actions to goal Actors are not “rational”March 2004 © Kathleen M. Carley – CASOS – ISRI - CMUGoal Based ApproachWorks reasonably well for things that can be monetarily characterizedGenerally produces behavior distinct from individual behaviorOptimization criteria need to be adjusted for cultures, individuals, etc. – so not that generalGets very complicated very quickly when: multi-criteria optimization is needed (multiple competing goals at multiple levels) goal change over time9March 2004 © Kathleen M. Carley – CASOS – ISRI -