CO600 Project Overview

Computing Laboratory, University of Kent
Proactivity using Bayesian
Methods and Learning
3rd UK-UbiNet Workshop, Bath
Lukas Sklenar
Organisation of the presentation
Collating Context
 Using Context

– Is it done? Limitations?
Bayesian Belief Networks
 Prediction & Proactivity

– Next steps
Collating Context

Large portions of current research
dedicated to collating context
– Particularly to achieve a high
confidence in the gathered data
– Reasoning under uncertainty, e.g.
inference has to be done on low-quality
sensor data
Collating Context - mechanisms

Many mechanism exist to help with
the interpretation of gathered context
– Bayesian Networks, Neural Nets,
Biologically inspired solutions, etc.

Toolkits exist that provide higher
level context information
– Create abstractions over sensors
– Give (almost) human readable results
Examples of Toolkits

Location Stack
– http://portolano.cs.washington.edu/projects/lo
cation/

PlaceLab
– http://placelab.org/

The Context Toolkit
– http://www.cs.berkeley.edu/~dey/context.html

An Architecture for Context Prediction
[Rene Mayrhofer, Pervasive 2004]
Limitations

Context is collected, displayed
– Little is actually done with it
– Although can be useful when displayed to
others

Some implementations allow for better
use, usually via if-then-else rules
– Such rules work, but can be cumbersome
– Usually have to be added/removed manually
– Such rules not resilient to change
Improvements
Need for intelligent proactivity
 Should comply with Weiser’s vision
of disappearing hardware (and
software!)
 For such functionality we need
devices that behave intelligently
 We propose to use Bayesian Belief
Networks to provide this intelligence

Bayesian Belief Networks

A Bayesian network is a compact,
graphical model of a probability
distribution [Pearl 1988].
– A directed acyclic graph which represents
direct influences among variables
– A set of conditional probability tables that
quantify the strengths of these influences
– Mathematically correct and repeatable
Technology : BBNs – overview1
Multiple parents
possible
Rain?
Sunny
Cloudy
Rainy
Rain
15
25
60
No Rain
70
20
10
P(F)
Forecast?
Rain?
Rain
No Rain
30
70
P(R)
Multiple parents
possible
Take Umbrella?
P(U)
Forecast?
YES
NO
Sunny
0
100
Cloudy
20
80
Rain
70
30
Technology : BBN’s – overview2
Example in Netica.
www.norsys.com
Technology : BBN’s – Summary
BBN's are trees which you can use to
predict P(state|other states)
 Structure and influences can be
learned from past data and/or
constructed by domain experts
 Used to interpret sensor data
 Could be used to proactively activate
features/alerts/etc.

FOR ME INFO...
http://www.norsys.com/belief.html
http://www.murrayc.com/learning/AI/bbn.shtml
BBN Uses

Already used when interpreting sensors
Sensor
Sensor
Interpretation layer
Data
BBN Proactivity
Sensor
Sensor
Sensor
Sensor
Interpretation layer
Interpretation layer
Data
Data
BBN-based Proactivity Mechanism
Same engine?
More features (power?) for a user
Adding Proactivity with BBNs

Eg. Add a
threshold of say
50. If >50,
recommend to
take umbrella
Add a threshold to
trigger events for
every combination
 Add a satisfaction
measure
 Adapt network or
threshold or both
according to
satisfaction
Potential

Having an intelligent proactivity
mechanism/enabler
– Could be learned from observing user usage
history
– Or created by a domain expert
– Complex relationships could be used as input
for an intelligent trigger
– These relationships would be resilient to
changes in your typical environment
– Whether to proactively activate something or
not could be calibrated with use
The End – thank you
Presented by Lukas Sklenar
http://www.cs.kent.ac.uk/people/rpg/
ls85/index.html
[email protected]
QUESTIONS?