180 likes | 215 Views
Context situations policy. Daniel Cutting, Aaron Quigley University of Sydney. Introduction. Daniel Cutting Ph.D. candidate at University of Sydney (Aaron Quigley supervisor, John Zic associate supervisor) Part of the Smart Internet CRC About half-way through Ph.D.
E N D
Context situations policy Daniel Cutting, Aaron Quigley University of Sydney
Introduction • Daniel Cutting • Ph.D. candidate at University of Sydney (Aaron Quigley supervisor, John Zic associate supervisor) • Part of the Smart Internet CRC • About half-way through Ph.D. • Thesis area: application collaboration in pervasive computing environments Daniel Cutting
Outline • Pervasive computing • Motivating scenario (art gallery) • Middleware • data distribution policies • Context spaces • Application to scenario • Discussion Daniel Cutting
Pervasive computing • Mobile devices (constrained, wireless) + fixed infrastructure (powerful, wireline) • Hypothesis: applications in PCEs can be improved using context • maximise availability of data • minimise battery usage and network traffic • constrained by user preferences • use context to aid data distribution Daniel Cutting
Art gallery scenario Bob was here. Gillian Edward Bob Cynthia Sunflowers, Van Gogh Bob was here.
Art gallery scenario • Guide publishes data that is pushed to students (marking image of painting) • Repository shared by group stores long-lived data (group photo) • Public infrastructure stores persistent data (painting images, guest book) Daniel Cutting
Middleware • Publish-subscribe: good for events • markings on painting image • Tuple spaces: good for data persistence • guest book, group repository • Build middleware that combines the two Daniel Cutting
Middleware distribution • Distributing/storing data is a problem • many devices, some small, wireless • may have powerful fixed infrastructure, but sometimes purely ad hoc networks • Middleware needs flexible data distribution and storage policy • Use context to aid this policy Daniel Cutting
Context • Sensed/inferred values from environment, network, devices, applications and users • e.g. beacons, bandwidth, storage capacity, usage patterns, preferences • Complex to base policy on raw context • interpose symbolic situations • context situations distribution policy Daniel Cutting
Context spaces • Treat context as n-dimensional space • Each dimension is type of context • e.g. [bandwidth, storage capacity] • sample context vector might be [high,low] • Specific situation vectors also exist (statically specified or learnt over time) • Find “nearest” situation vector to convert context vectors to situation Daniel Cutting
Zz z z Context spaces Daniel Cutting
Dynamic clustering • Don’t specify situation vectors • Cluster context vectors to automatically identify inherent situations • How should policy act if no situations exist until run-time? • Situations can shift over time to reflect changes to contextual sources Daniel Cutting
Scenario: context situations • Decentralised • each device determines own context • To build context space, designer identifies available context, e.g. • local power, bandwidth, storage • neighbours’ power, bandwidth, storage • size, priority, relevance, persistence of data • painting beacons, etc. Daniel Cutting
Scenario: context situations • Select context for dimensions • data importance I, persistence P, size S • context vector is of form [I,P,S] • For static space, specify situations • signature, photo, demonstration • e.g. photo [0.1,0.8,0.8] is when data is not very important, persistent and large (like a photograph) Daniel Cutting
Scenario: situations policy • A device putting data into the middleware system can: • store locally, broadcast, broadcast digest • Make distribution policy using situations • signature broadcast • photo digest • demonstration store Daniel Cutting
Gillian Edward Bob Cynthia Scenario: context policy Group photo at Sunflowers Group photo at Sunflowers Group photo at Sunflowers Nearest situation vector is photo photo digest Unimportant (0.2) Long-lived (0.7) Large size (0.9)
Discussion • Representing nominal and cyclic dimensions is troublesome • Can situations policy be automated in clustered context space? • Unknown values in context vectors could cause spurious results - project to lower dimensions? Daniel Cutting
Static classification • During design-time • manually specify situation vectors • During run-time • measure raw context • determine context vector • find nearest situation vector based on a metric such as Euclidean distance • space is not altered - essentially a lookup Daniel Cutting