Predicting missing person behaviour using Bayesian Networks

1
Predicting missing person behaviour using
Bayesian Networks

• Adam Golding
• (ardarm_at_optushome.com.au)
• Supervisors Charles Twardy, Kevin Korb

2
Aims of this presentation

• The SAR domain
• Goals of this project
• Missing Person Data Analysis
• Causal Model Construction
• The SARBayes Tool
• Problems
• Results
• Future Work

3
The SAR domain

• Search And Rescue
• Land (Mountain, Desert, Jungle)
• Sea
• Air
• Search Techniques
• Ground-pounding
• Air-based
• Sniffer Dogs
• Man-tracking

4
The SAR domain (contd)

• Conducting a Search
• Initial probability (POC) map
• Resource Allocation
• Personnel Management
• Financial Planning
• Subject Predictions
• The Search Subject
• Characteristics (age, sex, ht, wt, type)
• Survivability
• Mobility

5
Previous SAR Research

• POS POD x POC
• Equipment Technology (POD)
• Searching techniques (POD)
• Resource Allocation (POD)
• Lost Person Behaviour (POC)
• Psychology
• Case History
• Area History
• Software

6
Goals of this project

• The SARBayes Tool
• Search And Rescue using Bayesian networks
• Another vote in initial consensus
• Accepted and used in SAR community
• Incorporated into larger packages
• Analysis of Missing Person Data
• Derive types of missing person
• (Causal) Model Construction and Selection
• Prediction method for SARBayes
• Compare against historical models

7
(No Transcript)
8
(No Transcript)
9
Missing Person Data

• 350 US Virginia SAR Data Cases
• 55 Australian Cases (insufficient)
• Analysis
• MML classification (Snob)
• Find useful categories of missing people
• e.g hunter, hiker, Alzheimer's, child, etc
• Define sub-classes in type network node
• Incorporation
• Used to train network (Netica)

10
Causal Model Construction

• Why Causal?
• SAR is a classic example of RUU
• Causal relationship between SAR attributes
• Access to structure knowledge
• Rik Head, Emergency System Technologies
• Victoria Police
• Clues in literature
• Learn Models from Data (caMML)
• Ultimate model used in SARBaye

11
The SARBayes Tool

• Windows Application
• Interface to causal network
• Initial Prediction for starting search
• Input
• Known case details
• Output
• POC temperature map, Stat rings, Probability
  graphs, Case files

12

• Results

13
Data Analysis

• Analysis of theoretical type variable
• 4 main types of missing person
• Type 1 Child
• Type 2 Elderly/Alzeheimers
• Type 3 Normal Adult (hunter/hiker/etc)
• Type 4 Other Adult (psych/despondent/other)

14
Interesting Points

• Traditional Missing person types found to be
  superficial eg hiker, hunter
• Distance traveled found to be independent of type
• Unknowns almost always found in bush
• Child 86 likely to be found alive
• Normal Adults 96 likely to be found with 60
  chance of being found alive

15
Causal Models Construction

• 3 main approaches
• 1) Run data through caMML using traditional
  missing person types
• 2) Run data through caMML using Snobs
  discovered types
• 3) Expert knowledge elicited model (Mr Rik
  Head)

16
(No Transcript)
17
Interesting points

• Type and Age predict nothing but each other
• Model is useless for prediction, as no input
  variables influence output

18
(No Transcript)
19
Points to note

• Inverse of previous model
• Very strong association with type
• Another ugly structured model

20
Expert Elicited models

• Full theoretical model
• Type, Age, Sex, DistPLS, Health, Resource,
  Location.
• Each variable has between 3 and 9 states
• Model almost fully connected
• Lots of arcs Too many probabilities!

21
(No Transcript)
22

• Simplified Model Required
• Snobs types can be described by Age and Mental
  Status
• 3 variable model Age, Mental Status, DistPLS.

23
(No Transcript)
24
Model Assessment

• Measure predictive accuracy of model using
  information theoretic reward scheme
• Still pending

25
SARBayes Implementation

• Written in Visual C
• Uses Netica API for Bayesian Network
  functionality
• Written with expectations of larger package
  integration

26
Typical Program Flow

• Input Map
• BMP image format
• Click on known UTM co-ordinates
• Click on PLS
• Set initial POC grid
• N x N grid of cells
• Each cell has unique properties, hence a unique
  POC

27
(No Transcript)
28
Mark Specific Map Areas

• The Roads Tool
• Roads have different probability values,
  infferred from models
• A road through a cell divides the cell into 3 new
  areas, each with a different POC
• The Area Tool
• Marking of areas with differing terrain
• May affect many existing areas

29
(No Transcript)
30
Entering Case Data

• Eg subject age and category
• Case data may be entered continually as details
  are discovered
• SARBayes modifies evidence in network, then
  re-compiles the network for updated inferences

31
(No Transcript)
32
SARBayes Output

• Shaded map of POC values
• Instant visual representation of POCs
• Allows resource allocation decisions to be made
  instantly
• Stat ring overlays
• Show likely distances traveled, from network
  inference
• Bar Graphs
• Show probabilities of all query nodes
• Case Files
• Printable text files of all probabilities

33
(No Transcript)
34
(No Transcript)
35
(No Transcript)
36
Problems and Solutions

• Small amounts of Aussie data
• Missing data
• Modeled explicitly
• Allows inferences to be made when its known that
  details are missing (will never be known)
• Inaccurate data
• Compare with other data
• Check against historical records
• Sparse/useless data (hard to classify)
• Get more data
• Simplify (generalise) before analysis

37
What we now have

• Program capable of outputting useful inferences
  form missing person models
• Framework of data analysis and model construction
  processes
• Growing database of Australian missing person
  case files

38
Possible future extensions

• Build Australian Models, test against other SAR
  models
• Integration with larger SAR management packages
• extension to incorporate resource allocation
  algorithms
• Use current case results to update data in
  dynamic belief networks

39
Conclusion

• Practical and interesting project
• Real-world application
• Presentable final product - SARBayes
• Good cause!
• Websites
• http//www.csse.monash.edu.au/ctwardy/sar.html
• http//www.members.optushome.com.au/ardarm