BIB-VERSION:: CS-TR-v2.0
ID:: ncstrl.dartmouthcs//TR2004-491
ENTRY:: February 26, 2004
ORGANIZATION:: Dartmouth College, Computer Science
TITLE:: Evaluating next-cell predictors with extensive Wi-Fi mobility data
TYPE:: Technical Report (paper)
REVISION:: 1
AUTHOR:: Song, Libo
AUTHOR:: Kotz, David
AUTHOR:: Jain, Ravi
AUTHOR:: He, Xiaoning
DATE:: February 2004
RETRIEVAL:: For a paper copy, email <reports@cs.dartmouth.edu>
RETRIEVAL:: For a paper copy, write to
       Technical Report Librarian
       Department of Computer Science
       Dartmouth College
       6211 Sudikoff Laboratory
       Hanover, NH 03755-3510
       USA
RETRIEVAL:: PDF at http://www.cs.dartmouth.edu/reports/TR2004-491.pdf
ABSTRACT:: 
Location is an important feature for many applications, and wireless networks can better serve their clients by anticipating client mobility. As a result, many location predictors have been proposed in the literature, though few have been evaluated with empirical evidence. This paper reports on the results of the first extensive empirical evaluation of location predictors, using a two-year trace of the mobility patterns of over 6,000 users on Dartmouth's campus-wide Wi-Fi wireless network. We implemented and compared the prediction accuracy of several location predictors drawn from four major families of domain-independent predictors, namely Markov-based, compression-based, PPM, and SPM predictors. We found that low-order Markov predictors performed as well or better than the more complex and more space-consuming compression-based predictors. Predictors of both families fail to make a prediction when the recent context has not been previously seen. To overcome this drawback, we added a simple fallback feature to each predictor and found that it significantly enhanced its accuracy in exchange for modest effort. Thus the Order-2 Markov predictor with fallback was the best predictor we studied, obtaining a median accuracy of about 72\% for users with long trace lengths. We also investigated a simplification of the Markov predictors, where the prediction is based not on the most frequently seen context in the past, but the most recent, resulting in significant space and computational savings. We found that Markov predictors with this recency semantics can rival the accuracy of standard Markov predictors in some cases. Finally, we considered several seemingly obvious enhancements, such as smarter tie-breaking and aging of context information, and discovered that they had little effect on accuracy. The paper ends with a discussion and suggestions for further work. 
NOTE:: 
A shorter,preliminary version of this paper is to appear at IEEE Infocom, March 7-11, 2004.
END:: ncstrl.dartmouthcs//TR2004-491