MELT 2008September 19, 2008

Nattapong Swangmuang and Prashant Krishnamurthy

Graduate Program in Telecom/NetworkingUniversity of Pittsburgh, PA USA

On Clustering RSS On Clustering RSS FingerprintsFingerprints for Improving for Improving Scalability of Performance Prediction of Indoor Scalability of Performance Prediction of Indoor

Positioning SystemsPositioning Systems

Paper’s goals

• Enhance the model� for analyzing Wi-Fi location fingerprinting-based system using a proximity graph

• Study characteristics of fingerprint clusters and applying it to the performance modeling

Research Questions

• Can we reduce computational effort and make the model more scalable ?

• How much is its impact to the prediction of precision performance ?

Clustering Methodology

• Different clustering methods:

• median and K-mean

• Model each cluster separately

• Use measurement data in an office building environment

� IEEE PerCom 2008’s paper

Analytical Model: Precision as Probability

P{correct}

P{error}

comparison variable

chance of picking k over i

weights against neighbors of i

� Multi-location system: computing exact probabilty requires complex jointed probabilities and becomes prohibitive

– Solution: approximate prob. Given a MS at the i-th grid point, the prob. of selecting each fingerprint is derived

Fingerprint Clustering Example

-60 -55 -50 -45 -40-85

-80

-75

-70

SIS410 (dBm)

SIS

50

1 (

dB

m)

IS Building with 25 Location Fingerprints (K-MEAN cluster)

R~

1

R~

2

R~

3

R~

4

R~

5

R~

6

R~

7

R~

8

R~

9

R~

10

R~

11

R~

12

R~

13

R~

14

R~

15

R~

16

R~

17

R~

18 R~

19

R~

20

R~

21

R~

22

R~

23

R~

24

R~

25

KK--meanmean

0 0.5 1 1.5 2 2.5 30

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

Distance (meters)

Cu

mu

lati

ve

Pro

ba

bil

ity

of

Se

lec

tin

g L

oc

ati

on

Average Error Distance Distribution of 25 Locations

nonelim

elim-nocluster

elim-median

elim-kmean

9 fps

6 fps

8 fps

No significant difference in performance

No. of Operations Comparison

Median K-MeanOffline 8,284 10,789 31,875Online 27 29 50Offline 38,147 42,230 149,940Online 44 46 84

ClusteringNo ClusteringPhase

Scenario1

Scenario2

Conclusion

� Empirical study shows the model with fingerprint clustering maintain good performance

• No difference in the CDF of error distance curve

� With Clustering, the model becomes more scalable

• Save many operations required from the model without clustering

• Reduce # operations during both the offline and online phases

MELT 2008September 19, 2008

Nattapong Swangmuang and Prashant Krishnamurthy

Thank you

On Clustering RSS On Clustering RSS FingerprintsFingerprints for Improving for Improving Scalability of Performance Prediction of Indoor Scalability of Performance Prediction of Indoor

Positioning SystemsPositioning Systems