Towards Ad-hoc Situation Determination Graham Thomson, Paddy Nixon and Sotirios Terzis.

26
Towards Ad-hoc Situation Determination Graham Thomson, Paddy Nixon and Sotirios Terzis

Transcript of Towards Ad-hoc Situation Determination Graham Thomson, Paddy Nixon and Sotirios Terzis.

Towards Ad-hocSituation Determination

Graham Thomson, Paddy Nixon and Sotirios Terzis

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Introduction

PlaceLab has been successful in making location information freely available for use in experimental ubiquitous computing applications.

We envisage a need for tools that can deliver a much richer set of contextual information.

The high-level situation of the current environment is a key contextual element.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Scenario

Jane captures a ‘coffee break’ snapshot on her smartphone and instructs it that both calls and messages should be announced audibly when she is in that situation.

Historical snapshot of ‘formal meeting’ captured from context server.

In a partner company’s building her smartphone is unable to determine the situation. The local context server reveals it is ‘formal meeting’, and the previously associated behaviours are applied.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Drawbacks of the state of the art Require an environment expert. Reasoning mechanisms must be manually

constructed and maintained. Difficult to specify sensor to situation

correlations on a large scale. Situation specifications will suffer from the

subjective bias of the expert. Reasoning is performed by a single oracle

which cannot exploit non-public knowledge. Once programmed, situations are fixed.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Drawbacks of the state of the art Require an environment expert. Reasoning mechanisms must be manually

constructed and maintained. Difficult to specify sensor to situation

correlations on a large scale. Situation specifications will suffer from the

subjective bias of the expert. Reasoning is performed by a single oracle

which cannot exploit non-public knowledge. Once programmed, situations are fixed.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Representing the situation

Contextual information is captured as a relation between two instances of a class.

Ubiquitous computing environments are open - any number and variety of people, devices, and software may appear within them.

Contextual information they produce are therefore also open, as the instances of a relation are drawn from a potentially infinite set.

Reasoning about contextual information is then made difficult, as many machine learning techniques make strict demands on the structure and constraints of the data, e.g. the C4.5 algorithm.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Representing the situation

Text classification aims to automatically assign documents to a given set of categories.

Documents may exhibit no regular structure.

The set of known terms used in the documents may grow with each new document that is classified.

Easy to see similarities between the task of text classification and that of situation determination.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Situation as a document

Document Bag of terms. Situation Bag of relations. A relation is expanded to facilitate

reasoning at an abstract level.

Jane Robertworks with

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Situation as a document

Person

Person

Jane Robertworks with

Robertworks with

works with

Janeworks with

Person Person

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Situation as a document

To reason about snapshots, we must transform them into a representation suitable for machine learning algorithms.

Use vector space model - considers a document to be a vector in a multi-dimensional Euclidean space, with each axis corresponding to a term.

Extra time axis added to vector, scaled from 0 (midnight), to 1 (just before midnight).

Approach based on Support Vector Machines (SVM), which are currently the most accurate classifiers for text.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Determining the situation

+

+

+

+

+

+-

-

-

-

-

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Determining the situation

+

+

+

+

+

+-

-

-

-

-

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Determining the situation

+

+

+

+

+

+-

-

-

-

-

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Determining the situation

+

+

+

+

+

+-

-

-

-

-

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Determining the situation

+

+

+

+

+

+-

-

-

-

-

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Drawbacks of the state of the art Require an environment expert. Reasoning mechanisms must be manually

constructed and maintained. Difficult to specify sensor to situation

correlations on a large scale. Situation specifications will suffer from the

subjective bias of the expert. Reasoning is performed by a single oracle

which cannot exploit non-public knowledge. Once programmed, situations are fixed.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Process of interaction

Determining the situation within an environment is a cooperative effort between each participant (device) within it.

Each participant’s view is the union of its private, privileged, and public relation sets.

Make simplifying assumption that participants are in the same situation if they are in the same room.

Interaction proceeds as in the following five steps:

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Process of interaction

Stage 1 – Proposal of current situation

Formalmeetin

g

Formalmeetin

g

Formalmeetin

g

Formalmeetin

g

Coffeebreak

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Process of interaction

Stage 2 – Sharing of relevant snapshots

+

+++

+ +

++

+--

- --

-

-

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Process of interaction

Stage 3 – Correction (freshness, confidence, accuracy, derivation algorithm)

P = 2.3

P = 2.35

P = 1.8

P = -34P = ?

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Process of interaction

Stage 4 – Heuristic selection (strongest non-obvious indicator)

Formalmeetin

g

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Process of interaction

Stage 5 – Learning (dynamic adaptation to drift)

Formalmeeting-

+ --++

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Drawbacks of the state of the art Require an environment expert. Reasoning mechanisms must be manually

constructed and maintained. Difficult to specify sensor to situation

correlations on a large scale. Situation specifications will suffer from the

subjective bias of the expert. Reasoning is performed by a single oracle

which cannot exploit non-public knowledge. Once programmed, situations are fixed.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Preliminary results

Situations: Formal group meeting European project meeting Informal meeting Coffee break Private study Movie night

Ubicomp environments are highly dynamic - new situations will continually appear, while current situations will cease to recur.

The system must achieve an acceptable determination accuracy using as few snapshots as possible.

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Preliminary results

Average determination accuracy

0

10

20

30

40

50

60

70

80

90

100

0 20 40 60 80 100 120

Number of snapshots

Acc

ura

cy (

%)

Perv

asi

ve a

nd

Glo

bal C

om

pu

tin

g

www.smartlab.cis.strath.ac.uk

Preliminary results

Almost all misclassifications occurred between the three meeting situations.

THANK YOU!