1. Sense and Respond Systems for Crisis Management: The Event Web Mani Chandy Computer Science California Institute of Technology mani@cs.caltech.edu http://www.infospheres.caltech.edu/crisis_web/executive-summary.html Infospheres Group: Dan Zimmerman, Brian Aydemir, Elliott Karpilovsky, Andre Khorlin, Lu Tian, Jonathan White iSpheres Corp.: Eric Bax, Greg Billock, Lee Rosen, Ian Swett 1

2. OutlineRequirements for crisis management.Abstraction: The Event Web.Architecture of the Event WebTheoryPrototype.Dagstuhl Workshop on Crisis Management: Feb 2004 Work at Caltech and at iSpheres Corporationhttp://www.ispheres.comhttp://www.infospheres.caltech.edu/crisis_web/executive-summary.htmlSense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation2 3. The Hypotheses1. Information technology for crisis management indifferent application spaces (financial, healthcare,natural disasters, terrorism) have a common set ofproblems.Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 3 4. The Hypotheses1. Information technology for crisis management indifferent application spaces (financial, healthcare,natural disasters, terrorism) have a common set ofproblems. 2. A layer can be abstracted that represents thesecommon problems.Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 4 5. The Hypotheses1. Information technology for crisis management indifferent application spaces (financial, healthcare,natural disasters, terrorism) have a common set ofproblems. 2. A layer can be abstracted that represents thesecommon problems. 3. Computer science theory and implementationpractices can be applied to this layer.Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 5 6. The Hypotheses1. Information technology for crisis management indifferent application spaces (financial, healthcare,natural disasters, terrorism) have a common set ofproblems. 2. A layer can be abstracted that represents thesecommon problems. 3. Computer science theory and implementation practicescan be applied to this layer. 4. Software platforms for this abstraction can betailored quickly to map to different crisis situations. Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation6 7. We Cannot Plan for Every ContingencyCrisis unfolds in unexpected ways.Requirement: Configure sense and respond platform when crisis strikes; continuously reconfigure platform as crisis evolves.Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation7 8. Example Pilot radios from a plane that a passenger seems to have SARS symptoms. The plane crashes.Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 8 9. Dynamic Coalitions of Multiple Groups Manage CrisesGroups that manage crises are changing collections of institutions and individuals. The public, connected through the Internet and wireless, plays an increasingly important role.Requirement: Flexible dissemination network without central data but with dynamic security policies. Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation9 10. The Role of Time and Place The response of anorganization often dependson the history of a crisis, andnot merely on itsinstantaneous state. Locationplays an important role. Relevance: respond to statetrajectories and not merely to currentstate.Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation10 11. Example Time: Moving-point average of typhoid cases for 2 daywindow exceeds moving-point average over 1-monthwindow, and, supply of antibiotics below threshold T. Location Geographical information systems: Alert allvehicles moving towards a bridge. Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 11 12. Crisis Managers Need Attention MultipliersRespondSense Invoke distributed servicesDetect events acrossin real-timeextended environment in real-timeAnalyzeAggregate events across multiple sources Friendly forces in area. I am under attack Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 12 13. Architecture Source of raw data: e.g., sensor Event processor: EventdisseminationInfluence environment: Actuators Event stream directorySense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation13 14. ArchitectureNassau CountyNavyHospitalTWAEvent Nassau CountyCIAdissemination PoliceTravel agentJohn Q. Public Event stream directory Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation14 15. Central idea of architecture of Event Web Compositional.Extensible component library.Components can be added, deleted, modified during acrisis.Components are loosely coupled. Communication between components: Events. Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 15 16. Outline Requirements for crisis management.Abstraction: The Event Web.Architecture of the Event WebTheoryPrototype.Dagstuhl workshop on Crisis ManagementSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 16 17. Theory (time permitting)Semantics of events Error analysis Abstraction:The programming paradigm: when-then rules. Stepwise refinementglobal when-then rules to sets of sensors, actuators, and eventprocessors with local when-then rules. Event notations Delta dataflow algorithm and chip design Incremental computation to process event streams. Optimization: intelligent sensors, capacity limitations,Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation17 18. Types of eventsTemperature in MinneapolisAn agent B generates an event to change other agents estimates of the state of the environment. TimeModel used by agent B of daily variation on Nov. 3 Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation18 19. Events that changes estimates Actual An agent generates an Model band event to change other agents estimates of the state of the environment.TimeSense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 19 20. Events that change estimatesAgent generates an event to change knowledge or estimates of other agents.eventTimeSense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation20 21. Events that change estimates In this case, an event is encoded as a timestamp and a bit indicating whether the band eventwas crossed above or below Time Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation21 22. Types of events: periodic events Temperature in MinneapolisAn agent samples valuesaccording to a specified sampling strategyTime The model deals with the sampling strategy Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation22 23. Errors due to delay delayReceiving agents now estimation of state whenno signal is received.TimeSense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 23 24. Errors due to delaynowdelayReceiving agentsEstimated estimation of state when distributionno signal is received. of currenttemperature TimeSense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation24 25. Errors Errors due to:loss of eventsdelays clock drift erroneous monitoring of environment erroneous computation Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 25 26. WHEN f(global_state) THEN generate event World is a multidimensional system with a huge number of dimensions. Information about different dimensions represented by discrete events generated at different times.Generate event whenDanger point system enters this space t=3 t=5t=7t=9 Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation26 27. The Event Web implements When-Then Source of raw data,Actuators e.g., sensorEventDisseminationAppsNetworkSets of When -Then RulesEvent processorsSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 27 28. Stepwise Refinement Crisis manager specifies global when-then rules.Stepwise refinement to:Sensors that generate raw events for the when predicate.Event processors that execute local when-then rules.Actuators that implement the final then action.So that collectively the distributed system implements the globalwhen-then rules. Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 28 29. Delta dataflow for executing when-then Rajit Manohar, Cornell: building asynchronous VLSI chip for delta dataflowAt Least 2 Moving point average fAvg >2-min avg >Threshold 60 min avgSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation29 30. Notations for When Then Rules Alternatives Same as query for a database: a rule is a persistentquery. SELECT s WHERE bGeneral structure dealing with histories, geography,action sequences (e.g., time series, location sensitive,workflow representation) Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation30 31. Incremental Compilation: Map Tree to Existing Node Structure on Some Machine cc ccc cc T>0P>1PT Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 31 32. Incremental Computation f(x1, , xk, xn)f x1 xk xn Change in valuexkSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation32 33. Incremental Computation New value: f(x1, , xk, xn)f(x1, , xk, xn)f x1 xk xn Change in valuexkSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation33 34. Incremental ComputationIdeally, cost of computing f(x1, , xk, xn) from f(x1, , xk, xn) should be New value: f(x1, , xk, xn) independent of n. f(x1, , xk, xn)Examples: Threshold functions Moving-point computations f Boolean operators x1 xk xn Change in valuexkSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation34 35. Outline Requirements for crisis management.Abstraction: The Event Web.Architecture of the Event WebTheoryPrototype.Dagstuhl workshop on Crisis ManagementSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 35 36. Prototype and product Goal: build a platform for implementing when-then rules.Show how global when-then rules can be refinedautomatically, or systematically by engineers, to obtainimplementations. Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation36 37. Generic asynchronous node in Event Web TimerFrom event dissemination networkMessage-triggeredprogramEvent inboxesTo event disseminationnetworkSense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation37 38. Prototype and product Event schema:Caltech Infospheres prototype: XML documentiSpheres Corp product: Serialized Java object.Event disseminationInfospheres: explicit subscription to event generator by eventconsumeriSpheres prototype: subscription deduced from high-level notation.Event notationInfospheres: only provides compositional structureiSpheres: explicit notation for parametric and time series analysis. Sense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation38 39. Processing an Incoming MessageIncoming message Current state (XML doc)Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 39 40. Processing an Incoming MessageIncoming message Choose XSLT transformation based on incoming message and current state Current state (XML doc) New state(XML doc)Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 40 41. Processing an Incoming MessageIncoming message Choose XSLT transformation based on incoming message and current state Current state (XML doc) Sequence of outgoing Messages (XML docs)New state(XML doc)Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation 41 42. Architecture of event processor Incoming events queues Data structures shared by event queues and when- then rules engine Rules execution Sense and Respond Systems for Crisis Management3 November 2003, Agent Based Modeling and Simulation42 43. Related Work Events: David Luckham, iSpheres patent appsContent-based subscription: Gryphon, SiennaReal Time Business IntelligenceReal Time Business RulesActive databasesGartner: Roy Schulte on Event-Driven Architecture Download the iSpheres Sense and Respond platformfrom http://www.ispheres.comSense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation 43 44. Outline Requirements for crisis management.Abstraction: The Event Web.Architecture of the Event WebTheoryPrototype.Dagstuhl workshop on Crisis Management Work at Caltech and at iSpheres CorporationDownload product: http://www.ispheres.comhttp://www.infospheres.caltech.edu/crisis_web/executive-summary.htmlSense and Respond Systems for Crisis Management 3 November 2003, Agent Based Modeling and Simulation44