Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

20
Prediction-based Handoff in Prediction-based Handoff in Heterogeneous Networks Heterogeneous Networks Bill Phillips 09 March 2007

description

March 9, 2007Prediction-based Handoff in Heterogeneous Networks 3 Motivation Accurate mobile user location prediction facilitates: –Smoother session handoff –Improved network resource allocation –Better mobility management –Enhanced assignment of cells to location areas –More efficient paging –Decreased infrastructure costs No location predication: resources potentially reserved in each cell Location predication: resources reserved in only cell B AA B

Transcript of Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

Page 1: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

Prediction-based Handoff in Prediction-based Handoff in Heterogeneous NetworksHeterogeneous Networks

Bill Phillips09 March 2007

Page 2: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

2

OutlineOutline• General motivation• Dartmouth large-scale Wi-Fi study• Predictors

– Markov– LZ– PPM– SPM

• Discussion of Dartmouth results• References

Page 3: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

3

MotivationMotivation• Accurate mobile user location prediction

facilitates:– Smoother session handoff– Improved network resource allocation– Better mobility management– Enhanced assignment of cells to location areas– More efficient paging– Decreased infrastructure costs

No location predication: resources potentially reserved in each cell

Location predication: resources reserved in only cell B

A AB

Page 4: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

4

Dartmouth Wi-Fi StudyDartmouth Wi-Fi Study• Data collected from April 2001 – May 2003 from

Dartmouth Wi-Fi campus network• Over 6,000 WLAN users

– Average of 3,000 users active each day• 543 access points cover entire campus (interior

and exterior)• Recorded registered access point every 5-

seconds by MAC address– Not physical location of user– Not physical movement of user

• Analyzed 4 domain-independent predictors– Markov– Lempel-Ziv Algorithm (LZ)– Prediction by Partial Matching (PPM)– Sampled Pattern Matching (SPM)

Page 5: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

5

Dartmouth CampusDartmouth Campus• 200 acres• 161 buildings

– 82 residential– 32 academic– 6 library– 19

recreational– 22

administrative

• 802.11b WLAN • 543 APs• 81 Subnets

400 feet

Page 6: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

6

Test EnvironmentTest Environment• Wireless network included all offices, classrooms,

dorm rooms, and athletic/recreational facilities• All students required to have a computer, 70%

laptop.• Indoor AP range 150-350 feet• Dartmouth population: 5,500 students & 1,215

faculty• Residential – dominated all traffic• A few APs never received any traffic due to

remoteness, malfunction, or misconfiguration• Data collection: syslog events (3.5 million

records), SNMP polling (193 million records), and tcpdump sniffers (22 APs, all packet headers)

Page 7: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

7

Format of data collectionFormat of data collection• Domain-Independent measurements:

– Only changes in location are recorded– No time data– No geographic data, no association between access

points, mobile users, and geographic coordinates

Sample cell map & Sample location history:

L = gbdcbgcefbdbde(off)ecd

g

b

c

de

f

Page 8: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

8

Distribution of data collectedDistribution of data collected

• 50% of users had a trace length smaller than 500 points

• 50% of users had a trace length greater than 500 points

• 50% of users visited less than 25 APs

• 50% of users visited more than 25 APs

CDF: # users vs trace length CDF: # users vs # APs visited

X XHalf the WLAN users remained in a fairly small subset of the overall network

Page 9: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

9

Markov PredictorsMarkov Predictors• Order-k Markov predictors, O(k), assume that the next

location can be predicted using the last k locations.• Example: L = abacbcbcabab…

– O(1): P(a|b) = 1/2, P(b|b) = 0, P(c|b) = 1/2• O(1) Markov predictor would not know whether to choose a or c as

the next location– O(2): P(a|ab) = 1, P(b|ab) = 0, P(c|ab) = 0

• O(2) Markov predictor would choose a as the next location• Probabilities are represented by a transition probability

matrix, M. Using the O(1) example:

• After each prediction, history is incremented, and M is recalculated

3 10 4 41 102 21 2 03 3

aM b

c

a b cMoving from this state

into this state

Rows always sum to 1

Page 10: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

10

Lempel-Ziv (LZ) PredictorsLempel-Ziv (LZ) Predictors• Often used in text compression (UNIX compress,

WinZip)• Consider the text compression implementation first

• Initialize dictionary with length 1 blocks

• Encode entry with index• Add 1 to the length of the

previously encoded entry, and enter into dictionary

a b b a a b b a a b a b b a a a a b a a b b a0 1 1 0 2 4 2 6 5 5 7 3 0Dictionary

Index Entry Index Entry 0 a 7 baa 1 b 8 aba 2 ab 9 abba 3 bb 10 aaa 4 ba 11 aab 5 aa 12 baab 6 abb 13 bbarepea

t

Page 11: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

11

Lempel-Ziv (LZ) PredictorsLempel-Ziv (LZ) Predictors• Parses string into unique substrings from left to right• Each new substring differs in length by only 1 from a

previously listed substring• Example: L = abacbcbcaba

– Substrings: a, b, ac, bc, bca, ba• Substrings are commonly represented by a LZ tree that

includes measurement frequency.

• Predication occurs by using the relative frequency of substrings

Empty String

a:2 b:4

c:1 c:2

a:1

a:1

is an empty string

Page 12: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

12

Lempel-Ziv (LZ) PredictorsLempel-Ziv (LZ) Predictors• Prediction probability is dependent on the prefix

frequency• Example: L = abacbcbcaba …

– Substrings: a, b, ac, bc, bca, ba• Prefixes:

– a (2 times)– b (4 times)– bc (2 times)

• P(a|L) = 1/4, P(b|L) = 1/2, P(bc|L) = 1/4

Empty String

a:2 b:4

c:1 c:2

a:1

a:1

Page 13: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

13

Prediction by Partial Matching Prediction by Partial Matching (PPM)(PPM)• Used in text compression

• Similar to Markov predictors in that a variable subset, k, of past measurements is used to predict the next location

• Differs from Markov predictors in that only the last immediate k measurements are used for the prediction

• PPM constructs multiple O(k) Markov models for this subset of k measurements, and blends them into a singel prediction

• Example: L = … abdcada

O(6) PPM uses the last 6 measurements, and O(1 to 6) Markov predictors to determine that the next most like value is a

Page 14: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

14

Sampled Pattern Matching (SPM)Sampled Pattern Matching (SPM)• Sampled Pattern Matching• Similar to Markov predictors, but with a variable

order that depends on the longest common substring

• Example: acabcdcabcbabc– Longest substring = abc (length n)– The variable order O(k=/n) depends on where 0 <

<1– = 1/3, yields a O(1) Markov predictor– = 2/3, yields a O(2) Markov predictor

• This process is repeated for each user after each predication = Computationally intensive

• is a predetermined, fixed constant

Page 15: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

15

Tie BreakingTie Breaking• At times the prediction algorithm has equal probability of

choosing two different outcomes. The following are 3 tie breaking techniques:– First added: use the location that appears first in the location

history– Most recently added: (opposite of first added) use the location

that does not appear first in the location string– Most recent: use the location that was most recently visited by

the mobile user• The Dartmouth study showed no difference in prediction

between any of the 3 methods, so the first added technique was used

• L = acabcbab...a if using first addedc if using most recently addeda if using most recent

Predicted valueAssuming equal probability of predicting a or c

Page 16: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

16

Markov ResultsMarkov Results

• This figure shows that the accuracy of Markov prediction improved with larger samples (rightmost curve)

• O(2) was the best predictor (rightmost curve, above)

• O(2) out-performed O(3) and O(4)

• Among 543 access points, pattern lengths of 3 or 4 were not as common as 2

Most accurate for location histories greater than 1000 (35% of users)

Page 17: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

17

Markov fine-tuningMarkov fine-tuning• Investigated potential improvements to the O(2) Markov

predictor in the case that the algorithm could make no prediction (i.e. encountered a new location pair) – Fallback to O(1)– Fallback to most frequently used location– Time aided prediction (classrooms during the day/dining

halls, recreational facilities at night)

• O(2) with fallback to O(1) was the best performer

• Time aided prediction topic for future study

Page 18: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

18

Other PredictorsOther Predictors

• LZ performed well, but not as good as Markov O(2)

LZ PPM SPM

• PPM performed equally well at O(2) , O(3), O(4), and O(5), so number of previously considered locations did not seem to matter once greater than 2

• SPM performed marginally better at = 0.5, but not as well as Markov O(2) with O(1) fallback

• = 0.5 is a good balance between too much and too little information

Page 19: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

19

ConclusionConclusion• More computationally complex predictors yielded

results that were at best equivalent to the simpler Markov model, O(2) with O(1) fallback

• All predictors (Markov, LZ, PPM, SPM) faired poorly with movement histories under 100 (not enough samples)

• There was negligible improvement for trace lengths greater than 1000 locations

• Best predictor, Markov O(2) with O(1) fallback– Median prediction accuracy of 63% overall– Median prediction accuracy of 72% for location histories

exceeding 1000– Not known whether this is the optimal predictor for

longer traces, or once time dependent measurements are considered

Page 20: Prediction-based Handoff in Heterogeneous Networks Bill Phillips 09 March 2007.

March 9, 2007 Prediction-based Handoff in Heterogeneous Networks

20

ReferencesReferences• David Kotz, and Kobby Essien, “Analysis of a

Campus-wide Wireless Network,” MOBICOM ’02, Sep 23-26, 2002.

• Libo Song, David Kotz, Ravi Jain, and Xiaoning He, “Evaluating Next-Cell Predictors with Extensive Wi-Fi Mobility Data,” IEEE Transactions on Mobile Computing, vol. 5, no.12, pp. 1633-1649, Dec. 2006.

• Leon-Garcia, Alberto. Probability and Random Processes for Electrical Engineering, Addison Wesley, 1994.

• http://www.data-compression.com/lossless.html#lz