Upload
others
View
2
Download
0
Embed Size (px)
Citation preview
Bayesian networks for decision making under uncertaintyHow to combine data, evidence, opinion and guesstimates to make decisions
InformationTechnology
Professor Ann Nicholson
Deputy DeanFaculty of Information TechnologyMonash University
SourcesofUncertainty
• Solution:ProbabilityTheory
Ignorance PhysicalrandomnessComplexity
Thereasoningprocess1.Startwithabeliefina
proposition“Theimportedmangois
infestedwithapest”“Thepatienthaslung
cancer”“Theapplicantwillpayback
theloan”“Salesofaproductwill
increase”
“Beliefs”• Veryunlikely• 1%chance• Oddsof100to1• 0.01probabilityFrom• Gutfeeling• Expertopinion• Data
Thereasoningprocess1.Startwithabeliefina
proposition“Theimportedmangois
infestedwithapest”“Thepatienthaslung
cancer”“Theapplicantwillpayback
theloan”“Salesofaproductwill
increase”
2. New information becomes available
“Itisfromacountrywherethepestisendemic”
“Thepatientisasmokerandhasacough”
“Theapplicantdefaultedonaprevious loan”
“Somemodelsarerecalledforsafetyreasons”
3.Updateyourbeliefs But how?
Probabilitytheoryforrepresentinguncertainty
• Propositionsareeithertrueorfalse“Patienthascancer”
• Assignsanumericaldegreeofbeliefbetween0and1topropositionsP(“Patienthascancer”)=0.001prior probability (unconditional)
• WecannowrepresenttheimpactofevidenceonbeliefP(“Patienthascancer|positivemammogram”) =0.8• Conditional probability• Or,posterior probability(bywayofBayes’theorem)
TheBayesianapproach
TheRev.ThomasBayes1702?-1761
• Represent uncertaintybyprobabilities
• UseBayes’theorem:h=hypothesise=evidence
P(h|e) = P(e|h) × P(h)P(e)
Startingbelief=‘prior’
Newbelief
EstimatingRisk
Scenario: Anathleteistestedforsteroiduseandthetestcomesbackpositive.Youknowthat:• Onein100competitors arethoughttotake
steroids• Thetestisn’talwaysaccurate
Falsepositiverate:10%Falsenegativerate:20%
Q.Whatistheprobabilitytheathlete isadrugcheat?
Suppose:• h = “the athlete is taking steroids”• e = “test result is positive”And:• P(h) = 0.01 (one in 100 people)• P(e|h) = 0.8 (true positive rate)• P(e|not h) = 0.1 (false positive rate)
What is P(h|e)?
Bayes’TheoremforEstimatingRisk
Ingeneral,peoplecan’tdoBayesTheorem (well)offhand!
≈ 0.075 (7.5%)
And how do we scale up to X1, X2, ….. X100, …. X1000 ??
BayesianNetworks
• Developed bygraphicalmodeling&AIcommunities in1980sforprobabilisticreasoningunderuncertainty
• Manysynonyms– Bayesnets,Bayesianbeliefnetworks,directedacyclicgraphs,probabilisticnetworks
Judea Pearl 2012 Turing
Award
ProbabilisticGraphicalModels
…andquantifytheuncertainty
…capturetheprocessstructure(notblackbox)
Foratargetsystem…
Whyusemodels?• Increases understanding• Supportsdecisionmaking• Usenewdataandevaluationtoimproveovertime
NativeFishExample
Alocalriverwithtree-linedbanks isknowntocontainnativefishpopulations,whichneedtobeconserved.Partsoftheriver passthroughcroplands,andpartsaresusceptible todrought conditions.Pesticides areknowntobeusedonthecrops.Rainfall helpsnativefishpopulationsbymaintainingwaterflow,whichincreaseshabitatsuitabilityaswellasconnectivity betweendifferenthabitatareas.Howeverraincanalsowashpesticides thataredangerous tofishfromthecroplandsintotheriver.Thereisconcern thatthetrees andnativefishwillbeaffectedbydroughtconditionsandcroppesticides.Seehttp://bayesianintelligence.com/publications/TR2010_3_NativeFish.pdf
BayesianNetworks- Definition
Agraph inwhichthefollowingholds:
1. Asetofrandomvariables=nodes innetwork
2. Asetofdirectedarcs connectspairsofnodes
Structure represents the causal process.(Anything missing?)
BayesianNetworks- DefinitionAgraph inwhichthe
followingholds:1. Asetofrandom
variables=nodes innetwork
2. Asetofdirectedarcsconnectspairsofnodes
3. Eachnode hasaconditionalprobability table(CPT)thatquantifies theeffects theparentnodeshaveonthechild node
4. Itisadirectedacyclicgraph(DAG), i.e.nodirectedcycles
Each row sums to 1
WORST CASEBEST CASE
Beforeyouknowanything(noevidence)
(Screen shots from the Netica BN software)
Diagnosis
Effects
Causes
Prediction
Effects
Causes
Whatnext?
• Havemodel• Haveestimatesofposteriorprobabilities
Q.Howdoweusetheseprobabilities toinformdecisions (aboutactionorinterventions)?
RiskAssessment– thedecision-theoreticview
Risk=Likelihood xConsequence
P(Outcome|Action,Evidence) Utility(Outcome|Action)
Decision making is about reducing risk or “maximising expected utility”
Decision nodes
Utility nodes(cost/benefits)
Decisionnetwork=BN+decisions+utilities
Ourmethodology1:Buildamodel• E.g.Variables:patient’sdetails,
diseases, symptoms,interventions
• Costs/benefits:eg.$,QALY
Test
ReviewStructure
Parameters
Experts
DataLiterature
2:Embedmodelindecisionsupporttool• Diagnosis• Prognosis• Treatment• Riskassessment• Prevention
Design
Build
Review
Revise
Bayesiannetworksforfogforecasting(CollaborationwithAus Bur.ofMeteorology)
Phase1:(Boneh etal,2015)Inusebyweather forecastersinMelbourne since2006
Willtherebeafogbefore9amtomorrowmorning?
Bayesiannetworksforfogforecasting
Stage2:2013-15Researchproject(prototype)(Boneh etal.Inpreparation)• Explicittemporalmodelling,predictingtimeofonsetandclearance
CaseStudy:ModellingWillows(inSt.John’sRiverBasin,Florida)
Plant and seed collectionArtificial Islands
Transplanting
Changes in water depthMay 2010 Artificial Ponds
Greenhouse exptsGrazing
Evaluating germination
Experimental research program 2009-2011
DBNExample:ModellingWillows(inSt.John’sRiverBasin,Florida)(Wilkinsonetal.,2013)
Next state
State transitions
Current state
Scenarios
Yearling
Unoccupied
Sapling
Adult
Burnt Adult
Willow life-history stage transitions
cell size = 100m x 100m (1ha)CanalsCattailGrassSedgeMarshSawgrassHerbaceousMarshWillowSwampMixedShrubTreeIslandHardwoodSwampOpenWaterLevees
Architecture of the Integrated Management Tool
GIS data
ST-DOOBN:State-
TransitionDynamic OO
Bayesian Network
Seed Production & Dispersal Spatial OO Bayesian Network
(Chee et al., 2016)
Managing Complexity: Object-oriented Modeling
Enough_Bare_Ground
Burn_Decision
Burn_Intensity
Vegetation
Summer_PPTCanal_or_Center
Soil_Type
BurnEffect_on_Willow
Spring_PPTMech_Clearing
GrowingSeason_PPT
Available_Water_Spring Available_Water_Survival
Available_Water_GrowingSeasAvailable_Water_Germination
Level of Cover (T)
Stage (T)
Proportion_Germinating Seedling_Survival_Proportion
Level of Cover (T+1)
Stage (T+1)
Seed_Availability NumberGerminating NumberSurviving
Rooted_Basal_Stem_Diameter (T)
Rooted_Basal_Stem_Diameter (T+1)
Sapling_Transition
Yearling_Transition
UnOccupied_Transition
Burnt_Adult_Transition
NonInterv_Yearling_Transition
Adult_Transition
Seed_Production
Modelling Seed Production
SeedsPerStem = I * F * SwhereI is number of inflorescencesF is the number fruits per inflorescenceS is number of seeds per fruit
SeedsPerHA = SeedsPerStem * NumberOfStemsPerHA
ApplicationofBNsforcomplexenvironmentalmanagement:WesternGrasslandsReserves
(DSEProject2012-2013)(Sinclairetal.,InPreparation)
• 10,000hatoberestoredtonativegrasslands over10-20years
• Task:buildadynamicBNtoevaluate“what-if”scenariosover20years– arangeofmanagementstrategies
– foravarietyoflandtypes– explicitlyrepresentingcostsandenvironmentalvalues
BNsforRiskassessmentforLogExportsinNZ
CollaborationwithSCION(NZtimberresearch)
Potentia l s ink popula tion preva lence(s cript/BN 2.2.5)
Activity(BN 2.2.4)
Flight conditions(BN 2.2.2)
Deadwood (s imula tion 2.1.3)
Dis pers a l ke rne l(s imula tion 2.1.2)
Meteorology(GIS 1.1.1)
Fores try his tory(GIS 1.1.3)
Topography(GIS 1.1.2)
Pes t dis tribution(GIS 1.1.5)
Source popula tion preva lence(GIS 1.3.3)
Source popula tion preva lence (BN 2.2.3)
Potentia l s ink popula tion preva lence(GIS 1.3.5)
Tempera ture dependentdeve lopment
(s imula tion 2.1.1)
Tempera ture dependentdeve lopment(GIS 1.2.1)
Loca l popula tion preva lence(BN 2.2.6)
Loca l popula tion preva lence(GIS 1.3.6)
Flight condtions(GIS 1.3.2
Activity(GIS 1.3.4)
Deadwood(GIS 1.2.3)
Dis pers a l ke rne l(GIS 1.2.2)
Log pes t preva lence(Script 2.3.2)
Log pes t preva lence(DB 1.4.2)
Log ba tch pes t preva lence(Script 2.3.3)
Log ba tch pes t preva lence(DB 1.4.3)
Fores try his tory(log tag DB 1.1.6)
Fores t productivity(GIS 1.1.4)
Mature adults(BN 2.2.1)
Mature adults (GIS 1.3.1)
Pes t infes ta tion ra te(BN 2.3.1)
Pes t infes ta tion ra te(DB 1.4.1)
BayesWatch:BNsforanomalydetectionintracking (CollaborationwithDSTO)(Mascaro etal.,2014)
• Task:Detectanomalousbehaviour ofvessels, cars,pedestrians
• OriginallyusedAISdatafromvessels inSydneyHarbour
• ApplyBNlearningtobuildmodelsofbehaviour
• Combines TimeseriesBN+TrackSummary BN
• Usemetricstoassessanomalous tracks
Modellingbushfireprevention&suppression(Penmanetal.,2015)
(ScreenshotfromtheGeNIe BNsoftware)
MedicalRiskAssessment:HeartDisease(Nicholsonetal.,2008;Floresetal.,2011)
SimpleBNmodel:NZapplesImportRiskAssessment(Wintle etal.,2014).Exampleof‘what-if’reasoning
Australia’s assumptions NZ assumptions
NZApplesBN(Wintle etal.,2014)Explicithandlingofquantities(ofproduct)
(Screen shot from the AgenaRisk BN software)
Bayesianmodellingoverview
BNModels
DataMining(CaMML,
Chordalysis,Snob)
ExpertElicitation
Existingmodels
Evaluation
Decisionsupporttools
Advancedmodelling
(dynamic/timeseries,objectoriented)
Current research• FullOOBN
frameworkandmethodology
• Learningmodelswithunobservedvariables
• OnlineDelphi-basedexpertelicitation
• Visualisationofprobabilisticoutputs