Title: PREDICTING NETWORK AVAILABILITY USING USER CONTEXT

1
CanalAVIST NICTA, Australia
Title PREDICTING NETWORK AVAILABILITY USING
USER CONTEXT Speaker Upendra
Rathnayake Session Chair Dr Max Ott Time
1730 (AU)
CanalAVIST ICT Forum
2
Predicting Network Availability Using User
Context
Upendra Rathnayake , Max Ott UNSW, Sydney,
Australia NICTA, Australia
3
Outline

• Introduction and motivation
• Related work
• Modeling and experimenting details
• Analysis and results
• Prediction results
• Ranking variables
• Tuning model parameters
• Conclusion and future work

4
1. Overview

• Realizing broadband wireless services
  effectively over heterogeneous networks
• Drivers
• Multiple interfaces to access networks of varying
  capabilities
• Many BW hungry applications are NOT real-time
• Abundant on-board memory and processing power
• Idea Rush delivery when capacity is abundant
  Live off storage otherwise
• Less cost for user
• Less infrastructure cost for provider
• Enabler
• Predicting network availability

5
1. Diversity of RANs Applications

• Devices have multiple interface
• 3G, Wi-Fi, GPRS
• Different networks appear
• Cellular coverage probably always
• Wi-Fi like Hot Spots sometimes
• They have different characteristics
• Low bit rates cellular coverage
• Inexpensive bits in Hot Spots

• Conventional voice/text applications
• Regular calls, SMS
• Multimedia applications
• Real time
• Video conferencing, gaming
• Non real time
• Email
• Video on demand (progressive download)
• Software updates

6
1. Scenario VOD

• Downloading a Now in Wi-Fi
• video file going into
• watching congested 3G
  comes under another Wi-Fi
• If the mobile knows going under 3G shortly
• Mobile can buffer data and application can run on
  buffer
• Avoid using congested (costly) 3G bits to some
  extent
• Sometimes can totally avoid using 3G, and
  continue in next Wi-Fi

3G
7
1. Aim of research
Wi-Fi
Need to download some amount (S) of data within T
Dont use UMTS
Vertical handover
UMTS
Average rate r S/T
Done
Done
Now
Normal approach
Approach with predictions
8
2. Predicting network availability?

• People have regularities in behavior
• Daily commute of an office worker (regular
  places)
• Coming in the morning and going in the afternoon
  (regular times)
• Even taking the same path (regular paths)
• Accessing the same news web page everyday
  (regular activities)

9
2. State of the Art

• Domain independent models
• E.g. Context-for-Wireless Ahmad Rahmati, Rice
  University
• Predicting WLAN availability wrt GSM cell ID
• No of times WLAN seen on a particular cell
• Use few parameters
• Domain dependent models
• E.g. Network connectivity prediction Yves
  Vanrompay, Katholieke University Belgium
• User path prediction
• Networks are at locations
• Domain specific assumptions
• There are other parameters than location, which
  would
• Improve accuracy of predictions (Environment
  Context)
• E.g. Whether power is on AC or not, Time of the
  day, Acceleration etc
• We use any available context variables domain
  independently
• Context prediction
• E.g. An architecture - R. Mayrhofer, Johannes
  Kepler University Linz

10
3. A Context Prediction Approach

• Classifying inputs at a time and deriving states
• Predicting future states
• Interpreting the predicted states

11
3. Context state predictions

• Fits Markov models (of order n)
• Stay time of a state implicitly modeled
  (geometric)
• Semi-Markov models
• Stay time explicitly modeled
• Generalization of Markov models
• Can capture the sequence of activities
  irrespective of stay times
• Real state transfers
• We used Semi-Markov models
• Of order 1 and 2

12
3. Experiment

• Instrumented 4 mobile phones to log for every 30
  Sec
• Time of the day (morning/evening)
• WLAN AP availability
• LAN availability
• Power on AC or not
• Number of Bluetooth devices around
• GSM location area (LAC)
• Given to 4 users to use as their personal phone
• Two researchers, An HR person and an IT person
• Collected logs for over 4 weeks each

13
4. Prediction Results

• Predicted WLAN available minutes in next 5 minute
  blocks
• Actual probability ( actually available
  minutes/5)
• Predicted probability ( predicted available
  minutes/5)
• Predictions are accurate when user movements are
  less
• Mid of the day (mid of the graph)
• Averaged prediction errors in transit times
  (mornings evenings)
• Order 2 slightly over performs (less error than
  in order 1)
• Prediction error is less than 26 on average

14
4. Ranking Variables

• Collecting and using data needs processing/power
• Need to rank variables automatically according to
  their contributions to predictions
• Unimportant variables can be removed
• Can save power and processing
• Can adapt to each user

Importance - High
Importance - High
Importance - Less
15
4. Ranking Variables

• Used a conditional independence check
• Using above method, found GSM more important than
  Bluetooth information for example
• Prediction errors higher without GSM than without
  Bluetooth generally
• Also more the variables, less the error in
  predictions

16
4. Tuning model parameters

• E.g. Bluetooth variable (number of Bluetooth
  devices around the user)
• A binary variable - above/below some threshold
• How to set this threshold so that accuracy of
  predictions can be improved
• We used the same conditional independence check
• To find the relevance of Bluetooth variable at
  different cut off thresholds

• Prediction errors are less with optimal threshold
• Than threshold of 1 (for comparisons)

17
5. Conclusion Future Work

• Heterogeneous networks multi-interfaced devices
• There are non real time applications
• They can wait
• Make them wait if good networks to appear
• Or rush data transfer if going out of good
  networks
• Predicting Network Availability
• As a domain independent context prediction
  problem
• Learn important variables in predictions rank
  them
• Future work
• More user data, more sophisticated models for
  predictions
• An architecture for effective heterogeneous
  computing
• Taking user preferences, device concerns (power),
  predictions etc.
• To use heterogeneous networks effectively

18
Thank You
19
CanalAVIST NICTA, Sydney
End of Presentation by Upendra Rathnayake
CanalAVIST ICT Forum