Optimising Dynamic Composition of Bayesian Networks for Context Sensing and Inference

ocelotgiantAI and Robotics

Nov 7, 2013 (3 years and 5 months ago)

44 views

Optimising Dynamic Composition of Bayesian

Networks for Context Sensing and Inference


Korbinian Frank
*
°

and Matthias Röckl
*

*Institute of Communications and Navigation

German Aerospace Center (DLR)

Oberpfaffenhofen, Germany

korbinian.frank | matthias.roeckl@dlr.de

Tom Pfeifer
°

°
Telecommunications Software & Systems Group (TSSG)

Waterford Institute of Technology

Waterford, Ireland

t.pfeifer@computer.org

High level

Activity

Interaction

Situation

Location

Symbolic Location

Indoor Position

Outdoor Position

Weather

Temperature

Used Services

Service Context

Thermometer

Rain Sensor

Motion Activity

IMU

Availability

Agenda

Task list

Calendar

Requests

Time

Clock

Day of

the week

Season

Number

of Persons

In Vicinity

Noise Level

Ambient Microphone

Activity

of Person

Nearby

Breaking Bayesian Networks (BNs) for Context Inference from Sensor Networks into Bayeslets is a proven
approach for optimising performance in resource constraint ubiquitous computing environments. Automatic
selection and composition of Bayeslets faces the challenge that the related costs (inference time, memory
consumption) grow exponentially with the number of components. To minimise the BN to be evaluated
dynamically, this research investigates optimisation approaches which evaluate the added value of using a
particular Bayeslet versus its cost.

User A

High level

Activity

Availability

Agenda

Interaction

Situation

Time

Motion

Behaviour

Environment

Location

User B

High level

Activity

Availability

Agenda

Interaction

Situation

Time

Motion

Behaviour

Environment

Location

User C

High level

Activity

Availability

Agenda

Interaction

Situation

Time

Motion

Behaviour

Environment

Location

A Bayesian Context Inference Network (left) that is broken into Bayeslets (right)

Inference

should

only

consider

those

Bayeslets

that

significantly

add

value
.


As Bayeslets are heterogeneous, for every linkable candidate the following has to be determined:

How much new information can the linkable Bayeslet provide?

How costly is the integration?


Dynamic Composition of Bayeslets provides solutions for both problems by one of the following criteria:

Net Normalized Mutual Information:




Net Value of Information:





NetVoI

takes

into

account

all

available

information

at

request

time,

is

therefore

advantageous

over

NetNI

.

However

the

design

of

the

utility

function

to

reduce

uncertainty

in

the

target

random

variable

of

the

Bayeslet

is

challenging
.

Only

a

realistic

evaluation

will

finally

allow

us

to

rate

the

performance

of

both

criteria

and

to

select

the

best

one
.






















Term

Cost

evidence

given


actions

the

for
Utility

Expected

Maximum

Bayeslet

linked

a

from


evidence

external

and


evidence

given


actions

the

for
Utility

Expected

Maximum

Expected

)
|
(
)
|
(
)
(
)
(
;
|
)
|
(
,
,
e
A
A
e
i
e
e
C
A
MEU
A
MEU
P
NetVoI
e
e
i
e


















This

work

received

funding

from

the

EC’s

7
th

Framework

Programme

[FP
7
/
2007
-
2013
]

under

grant

agreement

no
.
215098

of

the

Persist

Collaborative

Project

and

from

the

Irish

HEA

through

the

PRTLI

cycle

4

project

“Serving

Society
:

Management

of

Future

Communication

Networks

and

Services”
.

),
1
1
(
|
|
log
)
)
(
)
|
(
(log
)
(
|
|
log
)
:
(
)
:
(
b
n
X
X
P
Y
X
P
E
Y
C
X
Y
X
I
Y
X
NetNI
a
X







X,Y

being Bayeslets,
a≥
1

constant