Professional Documents
Culture Documents
Learning ∗
Sinno Jialin Pan, James T. Kwok, Qiang Yang, and Jeffrey Junfeng Pan
Department of Computer Science and Engineering
Hong Kong University of Science and Technology, Hong Kong
{sinnopan,jamesk,qyang,panjf}@cse.ust.hk
10 10
Frequency
Ferris, Haehnel, & Fox 2006; Pan et al. 2006; Ferris, Fox, (a) Time Period 1 (b) Time Period 2
& Lawrence 2007). Figure 1: Variations of signal-strength histograms in two
However, many previous localization methods assume time periods at the same location from one access point
that RSS maps between the signal space and physical lo- In this paper, we extend the Manifold Co-Regularization
cation space is static, thus a RSS map learnt in one time framework to location estimation in a dynamic WiFi envi-
period can be applied for location estimate in latter time pe- ronment. We empirically demonstrate that this framework
riods directly without adaptation. In a complex indoor WiFi is effective in reducing the calibration effort when adapting
environment, however, the environment is dynamic in na- the learnt mapping function between time periods. This pa-
ture, caused by unpredictable movements of people, radio per contributes to machine learning by extending the Man-
interference and signal propagation. Thus, the distribution ifold Co-Regularization framework to a more general case,
of RSS values in training and application periods may be sig- in which the number of data in different views can be differ-
∗
Sinno J. Pan, Qiang Yang and Jeffrey J. Pan are supported by ent and there are only a subset of the data having pairwise
NEC China Lab (NECLC05/06.EG01). correspondence. We also contribute to location estimation in
Copyright ° c 2007, Association for the Advancement of Artificial pervasive computing, by developing a dynamic localization
Intelligence (www.aaai.org). All rights reserved. approach for adapting RSS mapping functions effectively.
Related Work tion, which maps the signal vectors to locations, f (1) learnt
(j)
in time period t1 and (3) l labeled data {(si , `i )}li=1 and
Existing approaches to RSS based localization fall into (j) l+uj
two main categories. The first approach uses radio prop- uj unlabeled data {si }i=l1 +1 collected in time period tj ,
agation models, which rely on the knowledge of access where `i is the location of ith reference point. Our objec-
point locations (Bahl, Balachandran, & Padmanabhan 2000; tive is to adapt the mapping function f (1) to a new mapping
LaMarca et al. 2005). In recent years, statistical machine function f , for each online time period tj .
learning methods have been applied to the localization prob-
lem. (Ferris, Haehnel, & Fox 2006) apply Gaussian process Overall Approach
models and (Nguyen, Jordan, & Sinopoli 2005) apply Ker- Our LeManCoR approach has the following assumptions:
nel methods for location estimation, respectively. In order 1. Physical locations → RSS Values: Physical locations that
to reduce the calibration effort, (Ferris, Fox, & Lawrence are close to each other should get similar RSS values.
2007) apply Gaussian-Process-Latent-Variable models (GP-
LVMs) to construct RSS mapping function under an unsu- 2. RSS Values → Physical locations: Similar RSS values at
pervised framework. (Pan et al. 2006) show how to apply corresponding physical locations should be close to each
manifold regularization to mobile-node tracking in sensor other.
networks for reducing calibration effort. 3. The RSS values could vary greatly over different time pe-
However, there are few previous works that consider dy- riods, while in the same time period they change a little.
namic environments with some exceptions. The LEASE
system (Krishnan et al. 2004) utilizes different hardware The first two assumptions have been proven to be mostly
systems to solve this problem. LEASE employs a number true in indoor localization problem in wireless and sensor
of stationary emitters and sniffers to obtain up-to-date RSS networks (Pan et al. 2006; Ferris, Fox, & Lawrence 2007),
values for updating the maps. The localization accuracy can over different time periods. For example, in Figure 2, RSS
only be guaranteed when these additional hardware systems SA should be more similar to RSS SB than RSS SC in the
0
are deployed in high density. To reduce the need of the ad- old signal space, and RSS SA should be more similar to RSS
0 0
ditional equipments, (Yin, Yang, & Ni 2005) apply a model SB than RSS SC in the new signal space, since location A is
tree based method, called LEMT, to adapt RSS maps by only closer to location B than location C in the physical location
0
using a few reference points, which are additional sensors space. The pair {SA , SA } is called RSS corresponding pair.
for recording RSS values over time. LEMT needs to build The third assumption is also often true in a real world.
A
a model tree at each location to capture the global relation- Y
calibrating new RSS samples at a few known locations and S2 SA 0 Location Space (2-dimension) X
S2
fitting a linear function between these values and the corre- SB
sponding values from the RSS map. 0
SB 0
SC
Problem Statement 0 SC
S1
Signal Space in Old Time Period (m-dimension)
0 S1
Signal Space in New Time Period(m-dimension)
*signal from two access points is shown here
Consider a two-dimensional localization problem in a dy- Legend signal vector grid point
namic indoor environment.1 A location can be represented signal boundary grid boundary
Pl β ∗ = (B2 − A2 A−1 −1 µ
( l1 A2 A−1 1
γI (1) (1) (2)
− f (2) (xi ))2 }. 1 B1 ) 1 Y1 + l2 Y2 ), (9)
l i=1 (f (xi )
(4) where
In Equation (4), the first three terms make the mapping func-
tion f (1) correctly classify the labeled examples and smooth B1 = γlI J10T J20 K2 ; A2 = γI 0T 0
l J2 J1 K1 ;
(1)
(1) γI
both in function space and intrinsic geometry space. The A1 = µ γI 0T 0
l1 J1 K1 + γA I1 + l J1 J1 K1 + l1 +u1 L1 K1 ;
next three terms make the same thing to the mapping func- (2) γI 0T 0
(2)
γI
1
tion f (2) . The last term make f (1) and f (2) agree the same B2 = l2 J2 K2 + γA I2 + l J2 J2 K2 + l2 +u2 L1 K1 .
labels, in our case locations, of the corresponding pairs. µ is We call this resulting algorithm extended Co-Manifold
a parameter to balance data fitting in the two views and γI is RLS (eManCoR). eManCoR extends the Co-LapRLS to the
a parameter that regularizes the pair mapping functions. situation that the number of training data in multi-view is
different and there are only a subset of the data having pair-
As mentioned above, we use graph Laplacian to approx- wise correspondence among different views. Thus, we can
(1)
imate the term kf kI . Thus the terms γI kf (1) k2I and apply eManCoR algorithm to learn a pair of mapping func-
(2)
γI kf (2) k2I in Equation 4 can be replaced with the follow- tions in dynamic localization problem. The localization ver-
γI
(1)
γI
(2) sion of eManCoR is called LeManCoR
(1)T
ing two terms l1 +u1 f L1 f (1) and l2 +u2 f
(2)T
L2 f (2) , re-
spectively. The LeManCoR Algorithm
Pl1 +u1 (1)
Using f (1)∗ = ∗
i=1 αi K1 (x
(1)
, xi ) and f (2)∗ = Our dynamic location-estimation algorithm LeManCoR,
Pl2 +u2 ∗ (2) (2) which is based on our Extension to Manifold Co-
i=1 βi K2 (x , xi ) to substitute f (1)∗ and f (2)∗ in
Equation 4, we arrive at the following objective function of Regularization above, has three phases: an offline training
the (l1 + u1 )-dimensional variable α = [α1 , · · ·, αl1 +u1 ] and phase, an online adaptation phase and an online localization
(l2 + u2 )-dimensional variable β = [β1 , · · ·, βl2 +u2 ]: phase.
• Offline Training Phase
(1) (1)
α∗ , β ∗ = argmin { 1. Collect l1 labeled signal-location pairs {(si , `i )}li=1
1
α∈Rl1 +u1 ,β∈Rl2 +u2 at various locations (the first l pairs are collected from
µ (1) reference points) and u1 unlabeled signal examples
l1 (Y1 − J1 K1 α)T (Y1 − J1 K1 α) + γA αT K1 α+ (1) 1 +u1
{sj }lj=l 1 +1
.
(1) (1)
γI T 1
− J2 K2 β)T (Y2 − J2 K2 β)+ 2. For each signal vector si ∈ S (1) , build an edge to its
l1 +u1 α K1 L1 K1 α + l2 (Y2
(2)
k nearest neighbors in Euclidean distance measure. After
(2)
γA β T K2 β +
γI T that, the adjacency graph and weight matrix can be used
l2 +u2 β K2 L2 K2 β+
to form the the graph Laplacian L1 .
γI 0 (1)
l (J1 K1 α − J20 K2 β)T (J10 K1 α − J20 K2 β)}, 3. Apply LapRLS to learn the mapping function fx and
(5) (1)
fy , which are for x and y coordinates, respectively.
where Ks is the kernel matrix over labeled and unlabeled
examples in each view. Ys is an (ls + us ) dimensional label • Online Adaptation Phase
vector given by: Y = [y1 , ..., yls , 0, ..., 0], where the first ls 1. For each time period tj , j ≥ 1:
examples are labeled while the others are unlabeled. In our (1) (1)
If j = 1, the mapping function fx and fy can be ap-
case, Ys corresponds to a vector of locations of each coordi- 0(j) (1)
nate. Js is a diagonal matrix given by Js (i, i) = |Ys (i)|. Js0 : plied to estimate the location directly, fx = fx and
0(j) (1)
is a (l × (us + ls )) given by Js0 (i, i) = 1 for i = 1, 2, ..., l, fy = fy .
otherwise, Js0 (i, j) = 0. l is the number of corresponding If j > 1, randomly collect a few unlabel data uj around
pairs. In the above formulas, s = 1 and 2, respectively. (j)
the environment and get l labeled data: {(si , `i )}li=1 ,
The derivatives of the objective function, respectively, where `i is the location of ith reference point, from ref-
(1) (j)
vanish at the pair of minimizers α and β: erence points. Thus, each pair {(si , si )} is a corre-
sponding pair, where i = 1, ..., l.
(1)
( lµ1 J1 K1 + γA I1 + γI 0T 0
(1)
γI 2. Apply eManCoR to learn the pair of mapping functions
l J1 J1 K1 + l1 +u1 L1 K1 )α ∗(1) ∗(j) ∗(1) ∗(j) ∗(1) ∗(1)
(6) {fx , fx } and {fy , fy }. We use fx and fy
− γlI J10T J20 K2 β = µ
l1 Y1, as the new mapping functions in time period tj instead of
∗(j) ∗(j)
fx and fy . This is because the data in tj are insuffi- We first show the comparison results in Figure 4 (see next
cient to build a strong classifier. The learning process of page) where the number of reference points is fixed to 10,
∗(j) ∗(j) ∗(1)
fx and fy is to adapt the parameter of fx and fy
∗(1) and we further show the impact of different number of ref-
0(j) ∗(1) 0(j) ∗(1) erence points to LeManCoR in Figure 5. In Figure 4, Le-
with new data in tj . Thus fx = fx and fy = fy . ManCoR denotes our model while LeMan denotes the sta-
• Online Localization Phase tic mapping function leant in time period 12:30am-01:30am
by LapRLS. LeMan2 gives that for each online time period,
1. For each time period tj , each signal vector s̃ collected by reconstructed mapping function using the up-to-date read-
0(j) 0(j)
a mobile node, ˜ ` = (x̃, ỹ) = (fx (s̃), fy (s̃)) is the ing of reference points and a few new unlabeled data by
location estimation. LapRLS. LeMan3 denotes constructing the mapping func-
2. Optionally, we can apply a Bayes Filter (Fox et al. 2003; tion from RSS data collected in 12:30am-01:30am, up-to-
Hu & Evans 2004) to smooth the trajectory and enhance date reading of reference points and a few unlabeled data.
the performance of localization. Figure 4(a) shows that when the training data and test data
are in the same time period, 12:30am-01:30am, LeMan can
Experimental Results get a high location estimation accuracy. However, when this
static mapping function is applied to different time periods,
We evaluate the performance of LeManCoR in an of-
the accuracy decreases fast. See Figure 4(b), 4(c), 4(d), 4(e),
fice environment in an academic building, which is about
4(f). The curves of LeMan2 and show that only several la-
30 × 40m2 , over six time periods: 12:30am-1:30am,
beled and a few unlabeled data are far insufficient to build a
08:30am-09:30am, 12:30pm-1:30pm, 04:30pm-05:30pm,
new classifier. In the contrary, our LeManCoR system can
08:30pm-09:30pm and 10:30pm-11:30pm. The distribu-
adapt the mapping function effectively by only 10 reference
tions of RSS values of three different time periods shown in
points and 300 unlabeled examples in new time period.
Figure 3. For every time period, we collect totally 1635 WiFi
examples in 81 grids, each of which is about 1.5 × 1.5m2 . Robustness to the Number of Reference Points
Thus we obtained six data sets of WiFi signal data. We In this experiment, we study how is the impact of the num-
install several USB Wireless network adapters on various ber of reference points to LeManCoR. We set the number
PCs around the environment. These adapters act as refer- of reference points from 5 to 20, and compare the accuracy
ence points to collect the labeled signal data over differ- with error distance 3.0m. The result is shown in Figure 5. As
ent time periods. We consider the midnight time period can be seen, LeManCoR is robust to the number of reference
12:30am-1:30am as the offline training phase, when we ran- points. 1
0.9
data, of which only 30% have labels, that means the cor- 0.8
responding locations are known. The other five periods are
considered as online phase. For each of them, we randomly 0.7
18
14
12
16
14
examples in online time periods, and observed that our Le-
ManCoR is also robust to the number of unlabeled data us-
16
12
14
Frequency
Frequency
Frequency
10
12 10
8 6
6
6
(a) 12:30-1:30 (am) (b) 8:30-9:30 (am) (c) 12:30-1:30 (pm) examples and 300 new signal examples only took 10 sec-
Figure 3: Variations of signal-strength histograms over dif- onds.
ferent time periods at the same location Conclusions and Future Work
Dynamic Location Estimation Accuracy In this paper, we describe LeManCoR, a dynamic localiza-
tion approach which transfers geometric knowledge of the
For evaluating the adaptability of LeManCoR, we compare WiFi data in different time periods by modified manifold co-
it with a static mapping method LeMan (Pan et al. 2006). regularization. Our model is based on the observations that
Since LeMan is also based on manifold setting, we can see in each different time period, similar signals strength im-
the adaptability of LeManCoR clearly by the comparison. ply close locations, and the pairwise signals over different
0.92 0.8 1
0.7 0.9
0.9
0.8
0.6
Culmulative Probability
Culmulative Probability
Culmulative Probability
0.88
0.7
0.5
0.86 0.6
0.4
0.5
0.84
0.3
0.4
0.82
0.2 0.3
Culmulative Probability
Culmulative Probability
0.5 0.6 0.6
time periods should correspond to the same locations. The Localization over Large-Scale 802.11 Wireless Networks.
mapping function between the signal space and the physi- In Proceedings of MobiCom 2004.
cal location space is adapted dynamically by several labeled Hu, L., and Evans, D. 2004. Localization for mobile sensor
data from reference points and a few unlabeled data in a new networks. In Proceedings of MobiCom 2004.
time period. Experimental results show that our model can Krishnan, P.; Krishnakumar, A. S.; Ju, W.-H.; Mallows,
adapt the static mapping function effectively and is robust to C.; and Ganu, S. 2004. A System for LEASE: Location
the number of reference points. In the future, we will ex- Estimation Assisted by Stationary Emitters for Indoor RF
tend LeManCoR to continuous time instead of discrete time Wireless Networks. In Proceedings of INFOCOM 2004.
periods.
LaMarca, A.; Hightower, J.; Smith, I.; and Consolvo, S.
References 2005. Self-Mapping in 802.11 Location Systems. In Pro-
ceedings of Ubicomp 2005.
Bahl, P.; Balachandran, A.; and Padmanabhan, V. 2000. Nguyen, X.; Jordan, M. I.; and Sinopoli, B. 2005. A kernel-
Enhancements to the RADAR user location and tracking based learning approach to ad hoc sensor network localiza-
system. Technical report, Microsoft Research. tion. ACM Transactions on Sensor Networks 1(1):134–152.
Belkin, M.; Niyogi, P.; and Sindhwani, V. 2005. On Pan, J. J.; Yang, Q.; Chang, H.; and Yeung, D.-Y. 2006. A
Manifold Regularization. In Robert G. Cowell and Zoubin Manifold Regularization Approach to Calibration Reduc-
Ghahramani., ed., Proceedings of AISTATS 2005. tion for Sensor-Network Based Tracking. In Proceedings
of AAAI 2006.
Ferris, B.; Fox, D.; and Lawrence, N. 2007. WiFi-SLAM Schölkopf, B., and Smola, A., eds. 2002. Learning with
Using Gaussian Process Latent Variable Models. In Pro- Kernels, Support Vector Machines, Regularization, Opti-
ceedings of IJCAI 2007. mization and Beyond. Cambridge, Massachusetts: MIT
Ferris, B.; Haehnel, D.; and Fox, D. 2006. Gaussian Press.
Processes for Signal Strength-Based Location Estimation. Sindhwani, V.; Niyogi, P.; and Belkin, M. 2005. A
In Proceedings of Robotics: Science and Systems 2006. Co-Regularization Approach to Semi-supervised Learning
Fox, D.; Hightower, J.; Liao, L.; Schulz, D.; and Borriello, with Multiple Views. In Workshop on Learning with Mul-
G. 2003. Bayesian filtering for location estimation. IEEE tiple Views at ICML 2005.
Pervasive Computing 2(3):24–33. Yin, J.; Yang, Q.; and Ni, L. 2005. Adaptive temporal
radio maps for indoor location estimation. In Proceedings
Haeberlen, A.; Flannery, E.; Ladd, A. M.; Rudys, A.; Wal-
of PerCom 2005.
lach, D. S.; and Kavraki, L. E. 2004. Practical Robust