SAJAL K. DAS CReWMaN LOCATION-AWARE RESOURCE MANAGEMENT IN SMART HOME ENVIRONMENTS Sajal K. Das, Director Center for Research in Wireless Mobility & Networking.

SAJAL K. DAS CReWMaN

LOCATION-AWARE RESOURCE MANAGEMENT IN SMART HOME

ENVIRONMENTS

Sajal K. Das, Director

Center for Research in Wireless Mobility & Networking (CReWMaN)

Department of Computer Science and Engineering (CSE@UTA)

The University of Texas at Arlington, USA

E-mail: [email protected]

URL: http://crewman.uta.edu

[Funded by US National Science Foundation]


What is a Smart Environment ?

• Saturated with computing and communication capabilities to make

intelligent decisions in an automated, context-aware manner

pervasive or ubiquitous computing vision.

• Technology transparently weaved into the fabric of our daily lives

technology that disappears. (Weiser 1991)

• Portable devices around users networked with body LANs, PANs

(personal area networks) and wireless sensors for reliable commun.

• Environment that takes care of itself or users intelligent

assistants provide proactive interaction with information Web.

Examples: Smart home, office, mall, hotel, hospital, park, airport



Smart/Pervasive HealthcareConsider a heart attack or an accident victimDesired actions

Coordinate with the ambulance, hospital, personal physician, relatives and friends, insurance, etc.

Control the traffic for smooth ambulance pass through Prepare the ER (Emergency Room) and the ER personnel Provide vital medical records to physician Allow the physician to be involved remotely …

On a Timely, Automated, Transparent basis

PICO (Pervasive Information Community Organization) http://www.cse.uta.edu/pico@cse

M. Kumar, S. K. Das, et al., “PICO: A Middleware Platform for Pervasive Computing,” IEEE Pervasive Computing, Vol. 2, No. 3, July-Sept 2003.

http://www.cse.uta.edu/pico@cse










Heart attack victim

Heart attack victim

Pervasive Healthcare

Ambulance

Ambulance

Victim-AmbulanceCommunity

Largercommunityto save patient

Physician

Hospital HospitalCardiacSurgeon

Nurse

• Spouse• Police• Traffic control • Insurance Co.


PICO Framework

Creates mission-oriented, dynamic computing communities of software agents that perform tasks on behalf of the users and devices autonomously over existing heterogeneous network infrastructures, including the Internet.

Provides transparent, automated services: what you want, when you want, where you want, and how you want.

Proposes community computing concept to provide continual, dynamic, automated and transparent services to users.


PICO Building BlocksCamileuns (Physical devices)

(Context-aware, mobile, intelligent, learned, ubiquitous nodes)

Computer-enabled devices: small wearable to supercomputers

Sensors, actuators, network elements Communication protocols

CamileunsAccess pointInternet Gateway Access point

Gateway

Bluetooth802.11bCellular…


PICO Building BlocksDelegents (Intelligent Delegates)

Intelligent SW agents and middleware Location/context-aware, goal-driven services Dynamic community of collaborating delegents Proxy-capable: exist on the networking infrastructure Resource discovery and migration strategies QoS (quality of service) management

Community

Delegents


Visitor’sDelegent

Camileuns + Delegents = Chameleons

Surveillance

Traffic Monitor

Information Kiosk

Police Community

Automobile Community

Streetlamp


PICO Architecture

PICO Middleware Services

Community

Delegents

CamileunsAccess point/Gateway Access point/

Gateway

Bluetooth802.11bCellular…


Smart Homes: Objectives Use smart and pro-active technology

Cognizant of inhabitant’s daily life and contexts Absence of inhabitant’s explicit awareness Learning and prediction as key components Pervasive communications and computing capability

Optimize overall cost of managing homes Minimize energy (utility) consumption Optimize operation of automated devices Maximize security

Provide inhabitants with sufficient comfort / productivity Reduction of inhabitant’s explicit activities Savings of inhabitant’s time

“The profound technologies are those which disappear” (Weiser, 1991)


Smart Home Prototypes /Projects

Aware Home (GA-Tech) – Determination of Indoor location and activities

Intelligent Home (Univ. Mass.) – Multi-agent systems technology for designing an intelligent home

Neural Network House (Univ. Colorado, Boulder) – Adaptive control of home environment (heating, lighting, ventilation)

House_n (MIT) – Building trans-generational, interactive, sustainable and adaptive environment to satisfy the needs of people of all age

Easy Living (Microsoft Research) – Computer vision for person-tracking and visual user interaction

Internet Home (CISCO) – Effects of Internet revolution in homes

Connected Family (Verizon) – Smart technologies for home-networking


MAVHome at CSE@UTA

MavHome: Managing an Adaptive Versatile Home

Unique project – focuses on the entire home

Creates an intelligent home that acts as a rational agent

Perceives the state of the home through sensors and acts on the environment through effectors (device controllers).

Optimizes goal functions: Maximize inhabitants’ comfort and productivity, Minimize house operation cost, Maximize security.

Able to reason about and adapt to its inhabitants to accurately route messages and multimedia information.

http://ranger.uta.edu/smarthome

S. K. Das, et al., “The Role of Prediction Algorithms in the MavHome Smart Home Architecture”, IEEE Wireless Communications, Vol. 9, No. 6, pp. 77– 84, Dec. 2002.


MavHome Vision

Face recognition, automated door entry

Smart sprinklers

Lighting control

Door/lock controllers,Surveillance system

Robot vacuum cleaner

Robot lawnmower

Intelligent appliancesClimate control

Intelligent Entertainment

Automated blinds

Remote site monitoring and controlAssistance for disabilities


MavHome: Bob Scenario 6:45 am: MavHome turns up heat to achieve optimal temperature for waking (learned)

7:00 am: Alarm rings, lights on in bed-room, coffee maker in the kitchen (prediction)

Bob steps into bathroom, turns on light: MavHome records this interaction (learning), displays morning news on bathroom video screen, and turns on shower (proactive)

While Bob shaves, MavHome senses he is 2 lbs overweight, adjusts his menu (reasoning and decision making)

When Bob finishes grooming, bathroom light turns off, kitchen light and menu/schedule display turns on, news program moves to the kitchen screen

(follow-me multimedia communication)

At breakfast, Bob notices the floor is dirty, requests janitor robot to clean house (reinforcement learning)

Bob leaves for office, MavHome secures the house and operates lawn sprinklers despite knowing 70% predicted chance of rain (over rule)

In the afternoon, MavHome places grocery order (automation)

When Bob returns, grocery order has arrived and hot tub is ready (just-in-time).


MAVHome: Multi-Disciplinary Research Project

Seamless collection and aggregation (fusion) of sensory data Active databases and monitoring Profiling, learning, data mining, automated decision making Learning and Prediction of inhabitant’s location and activity Wireless, mobile, and sensor networking Pervasive computing and communications Location- and context-aware middleware services Cooperating agents – MavHome agent design Multimedia communication for entertainment and security Robot assistance Web monitoring and control


MAVHome Agent Architecture Hierarchy of rational agents to meet inhabitant’s needs and optimize house goals

Four cooperating layers in an agent

Decision Layer

Select actions for the agent

Information Layer

Gathers, stores, generates knowledge for decision making

Communication Layer

Information routing between agents and users/external sources

Physical layer

Basic hardware in house

House Agent

Rooms/robots

Agent Agent Agent Network / mobile network …

Agent Agent Agent Network / mobile network …

Appliances/robots

Transducers/actuators

User Interface

External resources

Physical

• Sensors• Actuators• Networks• Agents

Communication

• Routing• Multimedia download

Information

• Data Mining• Action Prediction• Mobility Prediction• Active database

Decision

• MDP/policy•Reinforcement learning• Multiagent systems/ communication


Indoor Location Management Location Awareness

Location (current and future) is the most important context in any smart computing paradigm

Why Location Tracking ?

Intelligent triggering of active databases

Efficient operation of automated devices

Guarantees accurate time-frame of service delivery

Supports aggressive teleporting and location-aware multimedia services -- seamless follow of media along inhabitant’s route

Efficient resource usage by devices -- Energy consumption only along predicted locations and routes that the inhabitant is most likely to follow


Location Representation

Location Information

Geometric – Location information in explicit co-ordinates

Symbolic - Topology-relative location representation

Blessings of Symbolic Representation

Universal applicability in location tracking

Easy processing and storage

Development of a predictive framework


Indoor Location Tracking Systems

Research Prototypes Underlying Technology Location Data Granularity

Active Badge

(Univ. of Cambridge)

Infrared Symbolic Room-level

Active Bats

(Univ. of Cambridge)

Ultrasonic Geometric 9 cm

Cricket (MIT) RF and Ultrasound Symbolic 4 x 4 feet

RADAR (Microsoft) IEEE 802.11 WLANs Symbolic 3 – 4.3 m

Smart Floor

(Georgia Tech)

Pressure Sensors Geometric Position of sensors

Easy Living (Microsoft)

Vision Triangulation Symbolic variable

Motion Star Scene Analysis Geometric 1 m


Inhabitant’s Movement Profile Efficient Representation of Mobility Profile

In-building movement sampled as collection of sensory information

Symbolic domain helps in efficient representation of sensor-ids

Role of Text Compression Lempel Ziv type of text compression aids in efficient learning of

inhabitant’s mobility profiles (movement patterns)

Captures and processes sampled message in chunks and report in encoded (compressed) form

Idea: Delay the update if current string-segment is already in history (profile) – essentially a prefix matching technique using variable-to-fixed length encoding in a dictionary – minimizes entropy

Probability computation: Prediction by partial match (PPM) style blending method – start from the highest context and escape into lower contexts


MavHome Floor Plan and Mobility Profile

Bob’s movement profile: a j k k o o j h h a a j k k o o j a a j k k o o j a a j k k …

Incremental parsing results in phrases:

a, j, k, ko, o, jh, h, aa, jk, koo, ja, aj, kk, oo, jaa, jkk, ...

Sample Floor-plan Graph-Abstraction

Possible contexts: jk (order-2), j (order-1), (order-0)


Trie Representation and Phrase Frequencies

jk (order-2) j (order-1) (order-0)

k|jk (1)

|jk (1)

a|j (1)

aa|j (1)

k|j (1)

kk|j (1)

h|j (1)

|j (2)

a(4) aa(2) aj(1)

j(2) ja(1) jaa(1)

jk(1) jh(1) k(4)

ko(1) koo(1) kk(2)

o(4) oo(2) h(2)

Probability of jaa:

Absence in order-2 and order-1; escape probability in each order: ½

Probability of jaa in order-0: 1/30

Combined probability of phrase jaa :

(½) (½ )(1/30) = 0.0048

a (7) j (7) o (6)h (2) k (8)

j (1)a (2)

k (1)a (1)

h (1)a (2) k (2) k (2)o (2)

o (1)

o (2)

Phrases and frequencies of different orders

Phrases: a, j, k, ko, o, jh, h, aa, jk, koo, ja, aj, kk, oo, jaa, jkk, ...


=Probability Computation of Phrases

Probability of k

½ at the context of order-2

Escaping into next lower order (order-1) with probability: ½

Probability of k at the order-1 (context of “kk”): 1/(1+1) = ½

Probability of escape from order-1 to lowest order (order-0): ½

Probability of k at order-0 (context of ): 4 / 30

Combined probability of phrase k = ½ + ½ { ½ + ½ (4/30) } = 0.509

jk (order-2) j (order-1) (order-0)

k|jk (1)

|jk (1)

a|j (1)

aa|j (1)

k|j (1)

kk|j (1)

h|j (1)

|j (2)

a(4) aa(2) aj(1)

j(2) ja(1) jaa(1)

jk(1) jh(1) k(4)

ko(1) koo(1) kk(2)

o(4) oo(2) h(2)


Phrase Probabilities

0.0048

0.0048

0.0048

0.0048

0.0905

0.0809

0.0048

ja

jaa

jk

jh

a

aa

aj

0.5905

0.0809

0.0048

0.0048

0.0195

0.0095

0.0809

0.0095

k

kk

ko

koo

o

oo

h

j

Phrase Probability Phrase Probability

Probabilities of individual locations can be estimated by dividing the phrase probabilities into their constituent symbols according to symbol-frequency and adding up all such frequencies for a particular symbol (location)

Total probability for location k is:

0.5905 + 0.0809 + 0.0048/2 + 0.0048/3 = 0.6754

Bob’s movement profile: a j k k o o j h h a a j k k o o j a a j k k o o j a a j k k …


Probability Computation of Individual Locations

Location Probability

k

a

h

o

j

0.6754

0.1794

0.0833

0.0346

0.0207

Bob’s movement profile: a j k k o o j h h a a j k k o o j a a j k k o o j a a j k k

Phrases: a, j, k, ko, o, jh, h, aa, jk, koo, ja, aj, kk, oo, jaa, jkk, ...

Probabilistic prediction of locations (symbols) based on their ranking

Prime Advantages of Lempel-Ziv type compression – most likely location is predicted

Prediction starts from k and proceeds along a, h, o and j

a

j

hk

o


Characterizing Mobility from Information Theory

Movement history: A string “v1v2v3…” of symbols from alphabet Inhabitant mobility model: V = {Vi}, a (piece-wise) stationary,

ergodic stochastic process where Vi assumes values vi Stationarity: {Vi} is stationary if any of its subsequence is invariant

with respect to shifts in time-axis

Essentially the movement history “ v1, v2, …, vn” reaches the system as C(w1), C(w2), …, C(wn) where wi s are non-overlapping segments of history vi and C(wi)’s are their encoded forms

Minimizes H(X) and asymptotically outperforms any finite-order Markov model

The number of phrases is bounded by the relation:

nlnllnn vVvVvVvVvVvV ,...,,Pr,...,,Pr 22112211

)(log)()( xpxpXH

nn

nOnc

logloglog)(


Entropy Estimation Bob’s movement profile: a j k k o o j h h a a j k k o o j a a j k k o o j a a j k k …

For a particular depth d of an LZ trie, let H(Vi) represent entropy at ith level.

Running-average of overall entropy is: d

i ii VVVVHd

VH1 121 ,...,,|

1)(

a (7) j (7) o (6)h (2) k (8)

j (1)a (2)

k (1)a (1)

h (1)a (2) k (2) k (2)o (2)

o (1)

o (2)

5795.26

30lg

30

6

8

30lg

30

8

2

30lg

30

2

7

30lg

30

7

7

30lg

30

7)( 1

VH

9361.0

2

4log

4

2

2

4log

4

2

30

85log

5

1

2

5log

5

2

2

5log

5

2

30

73log

3

1

2

3log

3

2

30

7)|( 12

VVH

789.12

)|()()( 121

VVHVHVH


LeZi-Update: Location Prediction Scheme

Init dictionary, phrase w

loop

wait for next symbol v

if (w.v in dictionary)

w := w.v

else

encode <index(w), v>

add w.v to dictionary

w := null

forever

Initialize dictionary := empty

loop

wait for next codeword<i, s>

decode phrase := dictionary[i].s

add phrase to dictionary

increment frequency of every prefix

of every suffix of phrase

forever

EncoderDecoder

A paradigm shift from position based update to route based update

Encoder: Collects symbols and stores in the dictionary in a compressed form

Decoder: Decodes the encoded symbols and update phrase frequencies


Predictive Framework: Route Tracking

Probability of a set of route sequences depends exponentially on relative entropy between actual route-distribution and its type-class

Route-sequences away from actual distribution have exponentially smaller probabilities

Typical-Set – Set of sequences with very small relative entropy

Small subset of routes having a large probability mass that controls inhabitant’s movement behavior in the long run

Concept of Asymptotic Equipartition Property (AEP) helps capture inhabitant’s typical set of routes


Probability Computation of Typical Routes

From AEP, typical routes classified as: { : 2 -1.789 L() - Pr[]}

where L() is the length of phrase and is a very small value

Threshold-probability of inclusion of a phrase into typical-set

depends on its length L()

At our context: L() Threshold Probability

1 0.289

2 0.080

3 0.002


Capturing Typical Routes

0.0048

0.0048

0.0048

0.0048

0.0905

0.0809

0.0048

ja

jaa

jk

jh

a

aa

aj

0.5905

0.0809

0.0048

0.0048

0.0195

0.0095

0.0809

0.0095

k

kk

ko

koo

o

oo

h

j

Phrase Probability Phrase Probability

At this point of time and context, the inhabitant is most likely to move around the routes along Bedroom 2, Corridor, Dining room and Living room

Typical Set of route segments comprises of : { k, kk, koo, jaa, aa }


Bob’s Movement along Typical Routes

a

j

k

o

Typical Route: k o o k j a a

Bedroom 2, Corridor, Dining room and Living room


Energy Consumption Static Energy Plan

Devices remain on from morning until the inhabitant leaves for office and again after return at the end of the day.

Let Pi : power of ith device; M : maximum number of devices; t : device-usage time; p(t) : uniform PDF. Expected average energy consumption:

M

i i

b

a

b

a

M

i i

M

i istat Pab

dtab

tPdtttpPenergyE

111 2)(

Using typical values of power, number and usage-time for lights, air-conditioning and devices like television, music-system, coffee-maker from standard home, static energy plan yields ~ 12–13 KWH average daily energy consumption.

Worst-Case scenario


Energy Consumption Optimal (Manual) Energy Plan

Every device turned on and off manually during resident’s entrance and exit in a particular zone.

Pi,j : power of ith device in jth zone; : max # devices in a zone; R : # zones; t : device-usage time in a zone; p(t) : uniform PDF.

Expected average energy consumption:

n

j iji

q

p

n

j ijiopt P

pqdtttpPenergyE

1 1,

1 1, 2

)(

Using standard power usage, optimal energy plan results in ~ 2–2.5 KWH of average daily energy consumption.

Optimal Scenario

But lacks automation and needs constant manual intervention


Energy ConsumptionPredictive Energy Plan:

Devices turned on and off based on the prediction of resident’s typical routes and locations (Incorrect prediction incurs overhead)

Devices turned on in advance – existence of time lag (t)

s : predictive success-rate. As s 1,

E[energypredict] E[energyopt]

sPpq

energyEn

j iji

tpredict

1 1,2

For the scenario, predictive scheme yields ~3-4 KWH consumption

Successful prediction reduction of manual operations and saving of inhabitant’s invaluable time inhabitant’s comfort


Discrete Event Simulator

Simulation Structure

Event types: Daily actions of a user, e.g., sleeping, dining, cooking, etc.

Event Queue Priority Queue for buffering events

Events ranked according to time stamp.

Event Initializer

Generates the first event and pushes it into the event queue

Event Processing

Carried out with every event

Calls the event generator to generate next event and pushes it into the queue

Calls various action modules depending upon the type of event


Simulation: Assumptions

Simulation Duration: 70 days

Different life-styles at weekdays and weekends

Mobility initiated as the inhabitant wakes up in the morning and starts daily-routine

Inhabitant’s residence-time at every zone – uniformly distributed between a maximum and a minimum value

Negligible delay between sensory data acquisition and actuator activation

Prediction occurs while leaving every zone

In inhabitant’s absence, the house has minimal activity to conserve energy resources


Granularities of Prediction

Predicting next zone Inhabitant’s immediate next zone / location A coarse level movement pattern in different locations

Predicting typical routes / paths Inhabitant’s typical routes along with zones More granular indicating inhabitant’s movement patterns

Predicting next sensor Every next sensor predicted from current sensor Large number of predictions lead to system overhead

Predicting next device Predict every next device the inhabitant is going to use Details of inhabitant’s activities can be observed


A Snapshot of Simulation

Master bedroom

Closet

closet

Bedroom Bedroom

Restroom

Restroom Wash

room

kitchen

Living Room

kitchenkitchenkitchen

0

20

10

30

40

50

60

70

80

90

100

Success Rate

Corridor

kitchen

Dining RoomDining RoomDining RoomDining Room

4

2

6

8

10

12

14

Energy Savings

Static

Optimal

Predicted

Predicted Actual Correct Prediction

Dining Room

kitchen

GarageGarageGarage


Learning Curve and Predictive Accuracy

85% – 90% accuracy in predicting next sensor, zone and typical route

Route prediction accuracy slightly lower than location prediction, yet provides more fine-grained view about inhabitant’s movements

Only 4-5 days to be cognizant of inhabitant’s life-style and movements

Higher granularity keeps device prediction accuracy low (63%)


Memory Requirements

Variation of Success-rate with table-size

85% success rate with only 3–4 KB memory for inhabitant’s profile

Small size typical set (5.5% -- 11% of total routes) as typical routes


Energy Savings

Reduction in Average Energy Consumption

Energy along predicted routes / locations only – minimum wastage

Average energy consumption – 1.4 * (optimal / manual energy plan)

65% – 72% energy savings in comparison with current homes


Reduction in Manual Operations

Prediction accuracy reduction of manual operations of devices brings comfort and productivity, saves time

80% – 85% reduction in manual switching operations


Future Work

Route prediction and resource management in multi-inhabitant (possibly cooperative) homes

Design and analysis of location-aware wireless multimedia communication in smart homes

Integration of smart homes with wide area cellular networks (3G wireless) for complete mobility management solution

QoS routing in resource-poor wireless and sensor networks

Security and privacy issues


A. Roy, S. K. Das Bhaumik, A. Bhattacharya, K. Basu, D. Cook and S. K. Das, “Location Aware Resource Management in Smart Homes”, Proc. of IEEE Int’l Conf. on Pervasive Computing (PerCom), pp. 481-488, Mar 2003.

S. K. Das, D. J. Cook, A. Bhattacharya, E. Hierman, and T. Z. Lin, “The Role of Prediction Algorithms in the MavHome Smart Home Architecture”, IEEE Wireless Communications, Vol. 9, No. 6, pp. 77– 84, Dec. 2002.

A. Bhattacharya and S. K. Das, “LeZi-Update: An Information Theoretic Framework for Personal Mobility Tracking in PCS Networks”, ACM Journal on Wireless Networks, Vol. 8, No. 3, pp. 121-135, Mar-May 2002.

A. Bhattacharya and S. K. Das, “LeZi-Update: An Information Theoretic Approach to Track the Mobile Users in PCS Networks”, Proc. ACM Int’l. Conference on Mobile Computing and Networking (MobiCom’99), pp. 1-12, Aug 1999 (Best Paper Award).

Selected References


D. J. Cook and S. K. Das, Smart Environments: Algorithms, Protocols and Applications, John Wiley, to appear, 2004.

A. Bhattacharya, “A Predictive Framework for Personal Mobility Management in Wireless Infrastructure Networks”, Ph.D. Dissertation, CSE Dept, UTA (Best PhD Dissertation Award), May 2002.

A. Roy, “Location Aware Resource Optimization in Smart Homes”, MS Thesis, CSE Dept, UTA (Best MS Thesis Award), Aug 2002.

S. K. Das, A. Bhattacharya, A. Roy and A. Misra, “Managing Location in ‘Universal’ Location-Aware Computing”, in Handbook in Wireless Networks (Eds, B. Furht and M. Illyas), Chapter 17, CRC Press, June 2003.

Selected References


Technology Forecasts (?)

• ‘ Heavier-than air flying machines are not possible’ Lord Kelvin, 1895

• ‘I think there is a world market for maybe five computers’ IBM Chairman Thomas Watson, 1943

• ‘640,000 bytes of memory ought to be enough for anybody’ Bill Gates, 1981

• ‘The Internet will catastrophically collapse in 1996’ Robert Metcalfe

• ‘Long before the year 2000, the entire antiquated structure of college degrees, majors and credits will be a shambles’

Alvin Toffler


Concluding RemarksConcluding Remarks

““AA teacher teacher can never truly teach unless he is can never truly teach unless he is

still learning himself. A lamp can never light still learning himself. A lamp can never light

another lamp unless it continues to burn its another lamp unless it continues to burn its

own flame. The teacher who has come to the own flame. The teacher who has come to the

end of his subject, who has no living traffic end of his subject, who has no living traffic

with his knowledge but merely repeats his with his knowledge but merely repeats his

lesson to his students, can only load their lesson to his students, can only load their

minds, he cannot quicken them”.minds, he cannot quicken them”.

Rabindranath TagoreRabindranath Tagore (Nobel Laureate, (Nobel Laureate,

1913)1913)

SAJAL K. DAS CReWMaN LOCATION-AWARE RESOURCE MANAGEMENT IN SMART HOME ENVIRONMENTS Sajal K. Das, Director Center for Research in Wireless Mobility & Networking.

Documents

das crewman slide

das crewman pico framework

airport slide

community computing

ieee pervasive computing

smart home environments

b cellular slide

automated services