Wireless Intelligence: Wireless Technology meets Artificial Intelligence

1
Wireless Intelligence Wireless Technology meets
Artificial Intelligence

• Qiang Yang
• HKUST, Hong Kong
• qyang_at_cs.ust.hk
• http//www.cs.ust.hk/qyang

2
Forbes Magazine 2003 Article ltltDigitally
Monitoring Momgtgt

• Eric Dishman is making a cup of tea-and his
  kitchen knows it
• (at Proactive Health Research lab in Hillsboro,
  OR)
• tiny sensors monitor the researcher's every move.
  
• Radio frequency identification tags and magnetic
  sensors discreetly affixed to
• mugs, a tea jar, and a kettle,
• plus switches that tell when cabinet doors are
  open or closed,
• track each tea-making step.
• A nearby computer makes sense of these signals
• if Dishman pauses for too long, video clips on a
  television prompt him with what to do next.
• Why needed?

(courtesy CMU)
3
The Need

• Healthcare (as an example)
• in the United States alone, the number of people
  over age 65 is expected to hit 70 million by
  2030,
• doubling from 35 million in 2000
• In many countries, it follows the same trend
• Wireless Intelligence systems are beginning to
  appear
• that analyze changes in behavioral patterns over
  time to provide early warning of aging diseases
  such as Alzheimer's.

4
Many Application Scenarios

• Education/Tour guide
• M-commerce/Shopping Mall Advertisement
• Manufacturing/Logistic Control
• Retail Store Customer Support
• Hospital Aging/Patient Monitoring
• Critical Security Surveillance
• Emergency Rescue Service
• Nations Customs Booking and Tracking
• Military

5
The Opportunity Wireless sensors are affordable
now! (Ni 04)
Supercomputer
10,000,000
Mainframe
1,000,000
Server
100,000
Workstation
10,000
PC
1,000
PDA
100
10
Watch
Embedded systems
Smart Card
1
RF Chips
0.1
???
0.01
6
Levels of Intelligence

• First level location
• Where am I?
• Second level Context
• Whos around me?
• Whats around me?
• Whos in my group?
• Third Level movement aware
• Which destination/direction are you going?
• Are you taking a car or a bus?
• Fourth Level semantics
• Which are your plans and goals?
• What are your preferences?
• Do you need help?
• Are you abnormal?

7
Location Estimation Microsoft Research RADAR
(IEEE 802.11)
8
RADAR

• Deterministic approach based on NN
• Using the mean of the 20 samples at each location
  for each BS
• Result is quite good for 3 BS, but can be
  improved further!

9
The Radio Map

• Two phases in location estimation
• Offline training phase Pr( oj li )
• Online localization phase Pr( li o )
• Intensive manual labor in offline training of the
  radio map
• Also known as the sensor model

10
802.11b Signals a function of time

• Signals detected from one AP at HKUST, CS
  department (Yin and Chai, July 2004)

11
Probability Based methods

• k APs, Signal strength values set S (AP1s1,
  AP2s2, . , APksk)
• probability of all such S at a location during
  training ?Joint Probability Distribution P(Sx)
• Online find location with highest P(Sx).

12
Intelligent Location Estimation

• University of Maryland, College Park
• Joint Clustering based Method
• Rice University
• Robotics Based Method
• Hong Kong UST
• LEAPS

13
U. Maryland Joint Clustering

• WLAN Location Determination via Clustering and
  Probability Distributions
• Moustafa A. Youssef, Ashok Agrawala, A. Udaya
  Shankar, University of Maryland, 2003

• For k APs, joint prob dist is the distribution of
  
• P( AP1 s1,... ,APk sk x) at a location x
• Using Bayesian Independence assumption
• P(AP1s1,,APkskx) P(AP1s1x)P(APkskx)
• Cluster the radio map
• Intra-cluster Loc Estimation using ML

14
Robotics Based Method Rice U.
Online

• Robotics Based Location Sensing using Wireless
  Ethernet, Andrew M. Ladd, Kostas E. Bekris,
  Algis Rudys, Rice University, 2002

offline
15
Level I Location (Work at HKUST)

• LEAPS Intelligent AP Selection (Yang, et. Al)
• Yiqiang Chen (ICS, CAS Shanghai) and Jie Yin,
  Xiaoyong Chai, Qiang Yang, 2004

• Goal
• Suppose we have a large number of APs
• Which ones to select for location estimation?
• How do we optimally allocate APs for the best
  results?
• How do we save energy by looking up as few APs
  signals as possible?

16
Offline AP Selection

• First, gather data to build Radio Map
• At each location (grid point), record the signal
  strength values
• Take the average from 20 or more
• ltS1, S2, Sngt, where n is the total of APs in
  area
• Second, using feature selection for AP selection

17
Decision Tree for Each Cluster

• Intra-cluster Location estimation
• Same as Rules
• AP(1) High?
• AP(2)Low?
• LocationG1,
• Prob90
• Can further reduced APs used
• Uses Information Gain again

18
Online Location Estimation

• Consider the following example
• AP191, AP284, AP365, AP478
• Compute distance to each cluster
• Cluster 2 is determined, and the decision tree on
  the right is used G2 is the answer!

19
Analysis on Accuracy
20
Energy Savings
21
Level IV High-level User Behavior Recognition in
a Wireless LANAAAI 04
22
Setting in indoor environment

• Questions
• Where is the professor?
• What is he doing now?
• What is his ultimate goal?

Entrance2-to-Office
Stay-in-Office
Goto-Seminar1
Entrance1-Exit
23
Wireless LAN

• Use radio frequency (RF) based technique.
• RF signal strength can be collected from the base
  stations.

Base Station 3
Base Station 1
Time t ( 69
58
75 )
Time t1 ( 67
60
73 )
Base Station 2
24
General Framework
Goals
What is his ultimate goal?
Actions
What is he doing now?
Locations
Where is the professor?
Sensory Data
25
Sensor-to-Location Challenges

• RF signal strength is highly uncertain and noisy.
• Multi-path fading effects
• Environmental changes
• The location must be inferred accurately!

Goals
Actions
Locations
Sensory Data
26
Dynamic Bayesian Network

• Extension of Bayesian network
• T. Dean and K. Kanazawa, 1989 K. Murphy, 2002
• Graphical probabilistic model
• Sensor uncertainty
• Structured state space
• Stochastic process over time
• A coherent framework to manage all levels of
  abstraction

27
Applying Dynamic Bayesian Network
Action Model
Sensor Model
28
Sensor Model

• A model of the conditional probabilities
• A finite location space
• A finite observation space
• Location space ? a set of grid points on the map
• Observation
• Where is the average values of signals
  received from the base station .

29
A Two-level Model

• Sensor-to-Action Level a DBN model
• Action-to-Goal Level an N-gram model

30
An Example
G1 Go-to-Print-in-Room1 G2 Go-to-Seminar-in
Room2
A1
Signal vector lt58, 60, 45gt
lt56, 59, 48gt
P(G1) P(A1G1)P(D1A1) 0.5 P(G2)
P(A1G2)P(D1A1) 0.5
31
An Example
G1 Go-to-Print-in-Room1 G2 Go-to-Seminar-in
Room2
A1
A3
P(A2A1, L1) 0.5 P(A3A1, L1) 0.5
P(A2A1, L1) 0.4 P(A3A1, L1) 0.6
P(G1) P(A1G1)P(A3A1,G1) P(D1A1)P(D3A3)
0.5 P(G2) P(A1G2)P(A3A1,G2) P(D1A1)P(D3A3)
0.5
32
An Example
G1 Go-to-Print-in-Room1 G2 Go-to-Seminar-in
Room2
A1
A3
P(A6A3,G1) 0.1 P(A6A3,G2) 0.9
A6
P(G1) 0.1 P(G2) 0.9
The professor is pursuing G2.
33
Environment and Data Set

• The environment is modeled as a space of 99
  locations, each representing a 1.5-meter grid
  cell.
• 8 out of 25 base stations were selected.
• Data for sensor model
• At each state 100 samples were collected, one per
  second.
• Evaluation Data about 600 traces
• A professors 19 different activities.
• 3-fold cross-validation

99 locations 10 actions 19 goals
34
Evaluation Criteria

• Efficiency
• The average processing time for each observation
  in the on-line recognition.
• Accuracy
• The number of correct recognition divided by the
  total number of recognition.

35
Experimental Results

• Comparison about the average processing time for
  each observation

36
Experimental Results

• Comparison about average recognition accuracy
  w.r.t sampling interval

Sampling Interval 1s 1.5s 2s 2.5s 3s
Whole DBN 89.5 87.1 84.2 75.4 71.9
DBN Bigram 90.5 83.2 82.1 74.7 72.6
37
Related Work Learning and Inferring
Transportation Routines

• Infer locations of usual goal like home or work
  place.
• Infer mode of transportation
• Predict future movements (short and long-term)
• Infer flawed behavior or broken routine
• Robustly track and predict behavior even in the
  presence of total loss of GPS signal.

• Lin Liao, Dieter Fox and Henry Kautz (Univ of
  Washington, Seattle. AAAI04 Best Paper)
• Describe a system that creates a probabilistic
  model of a users daily movements through the
  community using unsupervised learning from raw
  GPS data.

38
Related Work F. Zhao et.Al, ISDQ sensor network
based tarcking
User query root
39
Ongoing Work Reducing Calibration Effort for
Location Estimation Using Unlabeled Samples

• Two phases in location estimation
• Offline training phase Pr( oj li )
• Online localization phase Pr( li o )
• Intensive manual labor in offline training
• Spatially and temporarily high density calibration

40
Illustration (Chai 2004)
Total amount of calibration effort Ns? Nl
41
M1Reducing the Sampling Time at Each Location

• Training datarange from 5 samples/locationto 60

85.1
73.0
62.8
42
M2 Reducing the Number of Locations Sampled
Interpolation

• la and lb are the locations sampled

• lc is the locations to be interpolated from
  sampled

Obtained from the training samples
d d1 d2
43
Modeling User Traces Using Hidden Markov Model
(HMM)

• An HMM is a quintuple ltL, O, ?, A, ?gt
• L location-state space l1 ,l2, ln
• O observation space o1 ,o2, om
• ? radio map Pr( oj li )
• A location-state transition Pr( lj li )
• ? initial state distribution Pr( li )
• HMM model parameter ? (?, A, ?)

44
Improve the Radio Map ? Using EM

• Let T denote the set of user traces
• The likelihood of obtaining T given the model
  parameter ? is
• Look for ? such that ? argmax?Pr(T ?)

45
Experimental Results M1 EM
Effect of varying the number of traces on
reducing the sampling time
46
Experimental Results M1 M2 EM
Accuracy improvement over sampled locations in L1
Accuracy improvement over all the locations (L
L1L2)
Accuracy improvement over interpolated locations
in L2
47
Ongoing Work Coping with the Dynamic Signal
Distribution Using Machine Learning

• Signals detected from one AP at HKUST, CS
  department (Yin Jie and Xiaoyong Chai, July 2004)

48
Conclusions

• Mobile computing environment presents
  unprecedented opportunities and challenges for
  intelligent computing
• Four levels Location, Context, Movement and
  Semantics
• Inherent issues
• Uncertainty
• Power efficiency
• Many hardware/software solutions exist
• Many applications needs to be explored
• A major challenge linking up the Web and the
  Wireless World