QoC: Quality of Context Improving the Performance of ContextAware Applications

1
QoC Quality of Context Improving the
Performance of Context-Aware Applications

• Tobias Zimmer

2
What is context?!

• (semantic) Context in Ubicomp
• Data that describes the environment, the
  situation
• Implicit input to applications
• Enhancement, simplification of user interfaces
• Reduction of explicit user interaction
• Basis for smart behavior of applications

• The term context is not well-defined
• Todays context models
• No separation of the processing of
  contextsemantic meaning vs. data structure
• Monolithic from application to networking

gt

• Interpreting context is highly complex
• High error rates
• Example ? recognition rate 84,26 Activity
  Recognition from User-Annotated Acceleration
  Data Ling Bao, Stephen Intille Proceedings of
  Pervasive Computing, 2004
• Little user acceptanceW. K. Edwards and R. E.
  Grinter, "At Home with Ubiquitous Computing
  Seven Challenges" in Proceedings of Ubicomp 2001

3
Goal

• Improving the context recognition rates in
  ubiquitous computing environments
• Focused on Large scale ubiquitous computing
  environments
• Highly dynamic
• Heterogeneity of applications and hardware
• Modularity and distribution of applications
• Mobility and restricted resources of the clients
• Development of a context management system for
  handling Quality of Context (QoC)

Assumption Context recognition in ubiquitous
computing environments can be improved, by
including context attributes into the
interpretation of context data.
4
Motivation AwareOffice
Office environment with active artifacts that
interact by exchanging context information

• Context producers and consumers
• Cups
• Chairs
• Whiteboard Pens
• Digital Camera
• Door Plate
• PDAs

5
Context Attributes

• Context attributes as basis for the development
  of a context management system
• Attributes basic properties of context
• Independence of semantic context models and
  context management
• Well defined service access points (SAP) for the
  semantic context model
• Information on context attributes as an
  additional input to semantic context models
• Assessment of QoC
• Efficient selection of context algorithms
• Reduction of the error rate in semantic decisions
  
• Context management
• Definition
• Provision
• Processing
• of general context attributes

• Context attributes
• Age
• Spatial origin
• Reliability
• Relation

6
Independence of Contexts
Artifact Type B
Artifact Type A
Artifact
Context Type b
Context
Context Type c
Context Type a
Communication Space
Artifact Type D
Artifact Type C

• Cycles
• Artifacts consume context that they were involved
  in producing
• Fan-out
• A large number of similar artifacts produce
  contexts that trace back to one single context
  source
• Artifacts misjudge the quality of context
• Solution Genetic Relation of Contexts (GRC)

7
Genetic Relation of Context Data

• Degree of relationship is used to filter
  non-independent contexts
• Creating a random genome G for each first order
  context
• Bit-vector length l
• Sequencing the Genome in single genes ?
• Every gene has a (functional) locus
• Stores information on the relationship
• System parameters of the genome
• Number of genes n
• Number of alleles r
• Derivation of higher order contexts
• Recombination of the parents genomes
• New crossover method Probabilistic Multi Site
  Crossover (PMSC)

8
Crossover in Genetic Algorithms

• Classic genetic algorithms use many different
  crossover methods
• Normally no sequencing of the genome
• Two parents produce two siblings by recombination
• Crossover should conserve schemata
• One Point Crossover / Two Point Crossover /
  K-point Crossover
• Random cutting sites
• Sequences of variable length
• Cut and Slice
• Variable length of the genome
• Uniform Crossover (UX) / Half Uniform Crossover
  (HUX)
• UX Bit-wise exchange with a given probability
• HUX exactly 50 of all differing bits are
  exchanged

9
Probabilistic Multi Site Crossover (PMSC)

• Based on sequencing a genome into single genes
• Conserves genes (sequences) in there locus, not
  schemata
• This preservers information on the degree
  relationship
• m parents produce one child
• Genes are handed down with an adaptable
  probability (corresponding to the information
  content)

10
Calculation of the Degree of Relationship

• Comparison of two genomes
• Genes at the same locus are compared
• Indicator function f()
• Calculation of the degree of relationship
• Ratio of matching genes and total length of the
  genome
• Relationship function rel()
• Resource conserving method!

11
Properties of GRC

• Allows to determine the interdependence of
  context data by estimating the grade of
  relationship
• No additional knowledge necessary
• Resource conserving algorithms for generating and
  analyzing the genomes
• Estimating the grade of relationship enables pre
  filtering of context data
• Good estimate for the real relationship
• Tends to over estimate
• Stable bound for the mean error
• Estimator for the first generation (2 parents)

12
Properties of the GRC

• Mean error in the estimation of the degree of
  relationship
• Depends on the number of alleles r and the number
  of parents m

13
GRC-System

• Decision on further processing of context data
  bases on simple threshold functions
• Mainly constant or linear functions in real world
  systems
• Standard configuration 100 genes, 256 alleles gt
  100byte genome
• First results from simulation scenarios
• Simulation of an AwareOffice environment
  with the Java Context Processing and
  Communication Simulator context_sim

14
Application of GRC First results

• Simulation of an AwareOffice environment with a
  potential cycle

5 cups6 chairs3 tables3 windows2 pens1
sponge1 camera1 door plate1 AC5
PDAs28 Artifacts
15
Application of GRC First results

• Simulation one PDA with deactivated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 3348
• Total number of contexts produced 1436
• Overall recognition rate 27.27
• Mean relatedness of consumed Cs 0.6698
• Var of relatedness of consumed Cs 0.22008
• --------------------------------------------
• Simulation one PDA activated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 1395
• Total number of contexts produced 465
• Overall recognition rate 72.92
• Mean relatedness of consumed Cs 0.0039
• Var of relatedness of consumed Cs 3.77E-5
• ---------------------------------------------

• Simulation one PDA with deactivated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 3348
• Total number of contexts produced 1436
• Overall recognition rate 27.27
• Mean relatedness of consumed Cs 0.6698
• Var of relatedness of consumed Cs 0.22008
• --------------------------------------------
• Simulation one PDA activated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 1395
• Total number of contexts produced 465
• Overall recognition rate 72.92
• Mean relatedness of consumed Cs 0.0039
• Var of relatedness of consumed Cs 3.77E-5
• ---------------------------------------------

• Simulation one PDA with deactivated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 3348
• Total number of contexts produced 1436
• Overall recognition rate 27.27
• Mean relatedness of consumed Cs 0.6698
• Var of relatedness of consumed Cs 0.22008
• --------------------------------------------
• Simulation one PDA activated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 1395
• Total number of contexts produced 465
• Overall recognition rate 72.92
• Mean relatedness of consumed Cs 0.0039
• Var of relatedness of consumed Cs 3.77E-5
• ---------------------------------------------

• Simulation one PDA with deactivated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 3348
• Total number of contexts produced 1436
• Overall recognition rate 27.27
• Mean relatedness of consumed Cs 0.6698
• Var of relatedness of consumed Cs 0.22008
• --------------------------------------------
• Simulation one PDA activated GRC
• --------------------------------------------
• Total active steps 1000
• Total number of contexts consumed 1395
• Total number of contexts produced 465
• Overall recognition rate 72.92
• Mean relatedness of consumed Cs 0.0039
• Var of relatedness of consumed Cs 3.77E-5
• ---------------------------------------------

16
Summery

• Goal Improving context recognition in complex
  ubiqutous computing environments
• Realized by introducing context attributes
• GRC new method to solve typical problem in this
  application domain
• Independent from semantic context information
• Introduction to functionality and implementation
  of GRC
• First results The relationship attribute
  provided by GRC can improve the recognition rate
  of a simulated artifact up to 45.

17
(No Transcript)
18
Summery
19
(No Transcript)
20
Properties of the GRC-System
21
GRC-System Examples
Step-by-Step
22
GRC-System Examples
Multiple
23
Qualität von Kontexten (QoC)

• Die QoC muss für jeden Kontext von der
  verarbeitenden Anwendung evaluiert werden
• Faktoren, die QoC beeinflussen können
• Alter der Daten
• Räumliche Distanz zur Kontextquelle
• (Abstraktions-) Grad des Kontextes
• Genauigkeit verarbeiteter Sensorwerte
• Zuverlässigkeit von Kontextalgorithmen
• Systematische Fehler
• Statistische Fehler
• Auswahl möglicher Ausgangsdaten
• Evaluierung der QoC kann jetzt durch Betrachtung
  der Kontextattribute erfolgen
• Durch einheitliche Modellierung der Attribute
  Vergleichbarkeit der QoC innerhalb einer Anwendung

24
Auswahl der Kontextattribute

• Analyse von Anwendungsszenarien in der Literatur
• Zwei Studien Szenarien und deren
  anwendungsorientierte Umsetzung in Ubicomp
• Gestaltung von Anwendungsszenarien
• T. Zimmer et al. "Leben in einer smarten
  Umgebung - Ubiquitous Computing Szenarien und
  Auswirkungen", Gottlieb Daimler- und Karl
  Benz-Stiftung, December 2003.
• Eigene Erfahrungen bei der Umsetzung von
  Ubicomp-Anwendungen
• T. Zimmer, Towards a Better Understanding of
  Context Attributes, PerCom 2004, IEEE
• AwareOffice
• Kontextsensitive Anwendungen im Büroumfeld
  (erste AwarePen Umsetzung)
• Basic preprocessing, Neural Networks and Fuzzy
  Logic algorithms in the TecO-Smart-it platform
  (erster AwarePen auf Particle Toolboximplementier
  ung)
• Artefakte für das Aware Office (Integration
  AwarePen, AwareSponge, AwareCam, RoomController,
  DoorPlate, ...)

25
Auswahl der Kontextattribute

• Alter und Räumliche Herkunft in der Liteatur
• A. Schmidt, M. Beigl, and H.-W. Gellersen,
  "There is more to context than location",
  Computers and Graphics Journal, vol. 33, no. 6,
  pp. 893 902, 1999
• A. Schmidt, "Ubiquitous Computing Computing in
  Context", Ph.D. dissertation, Lancaster
  University, Nov. 2002
• Verlässlichkeit in der Literatur
• A. K. Dey, J. Mankoff, G. D. Abowd, and S.
  Carter, "Distributed Mediation of Ambiguous
  Context in Aware Environments", in Proceedings of
  UIST 2002
• W. K. Edwards and R. E. Grinter, "At Home with
  Ubiquitous Computing Seven Challenges" in
  Proceedings of Ubicomp 2001
• Verwandtschaft
• Keine Verweise in der Literatur
• Probleminduziert
• Auffächerung Typisch für Multiuserszenarien in
  homogenen Umgebungen
• Ringschluss Gossiping-Anwendungen, Random
  Encounter, UC-Spiele
• Effiziente Selektion von Kontexten Modulare
  Anwendungen, mobile Clients

26
Lösung genetische Verwandtschaft

• Algorithmus in Pseudocode
• ltbegin generate genegt
• for each parent context c
• c.gene rand_Bitvector(n)
• return c.gene
• ltendgt
• ltbegin crossovergt
• for each parent context c
• c.genei sequence(c.gene, s)
• child.gene new gene()
• for i 0 to n/s
• child.genei rand_select(c.genei)
• return child.gene
• ltendgt

ltbegin get relatedness(c1, c2)gt
relation_counter 0 for i 0 to n/s
if (c1.genei c2.genei)
relation_counter relatedness
relation_counters/n ltendgt
27
Eigenschaften der GRC

• Ermöglicht die Bestimmung von Abhängigkeiten
  zwischen Kontextdaten durch die Schätzung der
  Verwandtschaft
• Kein zusätzliches Wissen nötig
• Ressourcen schonende Algorithmen für Generierung
  und Auswertung der Genome
• Schätzung der Verwandtschaft ermöglich das
  Filtern von Kontextdaten
• Guter Schätzer für die tatsächliche
  Verwandtschaft
• Tendiert zur Überschätzung
• Stabile Schranken für denmittleren Fehler
• Standard-Szenarien zur modularen Evaluierung
• Stufenweise Vererbung
• Mehrfachvererbung

Stufenweise Vererbung
Mehrfachvererbung
28
Artefakt
Kontext
Kommunikationsraum
Artefakt Typ B
Kontext Typ c
Kontext Typ a
Kontext Typ b
Artefakt Typ A
29
Stufenweise Vererbung
Mehrfachvererbung
Artefakt Typ B
Artefakt Typ A
Kontext Typ b
Kontext Typ c
Kontext Typ a
Artefakt Typ D
Artefakt Typ C