Combining reactive & deliberative agents for Combining reactive & deliberative agents for complete ecosystems in infospherescomplete ecosystems in infospheres

Fabien L. GandonFabien L. Gandon

Carnegie Mellon UniversityCarnegie Mellon University

“The diversity of resources and information in real infospheres calls for artificial ecosystems with a diversity of interacting agents ranging from reactive to deliberative paradigms and

maintaining the information ecology.” proceedings IAT p. 297

Overview of the talkOverview of the talk

General positionGeneral position

Argue for complete ecosystems in infospheresArgue for complete ecosystems in infospheres

Technological stanceTechnological stance

Summarize interests of XML for infospheresSummarize interests of XML for infospheres

Initial and current workInitial and current work

Examples of implementations and experimentsExamples of implementations and experiments

World World ‘ ‘ Wild ’Wild ’ Web Web Information resources and services:Information resources and services:

Situated and Situated and scatteredscattered on the net; on the net; Ever Ever changingchanging in form and content; in form and content; Ever Ever growinggrowing in size and in size and heterogeneityheterogeneity..

Overwhelming Overwhelming complexity for humanscomplexity for humans.. Overwhelming Overwhelming complexity for machinescomplexity for machines::

unorganized and too heterogeneous in form, unorganized and too heterogeneous in form, content, quality, content, quality, etc.etc. for direct automation. for direct automation.

Alike our own world:Alike our own world: Vast, distributed, heterogeneous landscape;Vast, distributed, heterogeneous landscape; Rich fertile soil of information resources;Rich fertile soil of information resources; Actors and resources are situated;Actors and resources are situated; Actors can perceive, act and interact.Actors can perceive, act and interact.

The infosphere metaphor The infosphere metaphor

Infosphere: the equivalent in information worlds of Infosphere: the equivalent in information worlds of our biosphere and its ecosystem.our biosphere and its ecosystem.

Ecosystem:Ecosystem: System encompassing System encompassing beingsbeings and and environmentenvironment;; Self-regulatingSelf-regulating through complex cycles involving through complex cycles involving

multiple types of interaction;multiple types of interaction; InteractionsInteractions between a huge variety between a huge variety of beingsof beings;; InteractionsInteractions beingsbeings and huge variety and huge variety environmentsenvironments..

Good news… convergence between:Good news… convergence between: Distributed AI / Multi-agent systems Distributed AI / Multi-agent systems (beings)(beings) Structured and semantic Web & Services Structured and semantic Web & Services (enviro.)(enviro.)

Bad news…Bad news…

Compartmentalized current trendsCompartmentalized current trends

Autecology and compartmentalized schoolsAutecology and compartmentalized schools Focus on Focus on one type of agentone type of agent

e.g. information extraction.e.g. information extraction.

Follow one Follow one school of thoughtsschool of thoughts

e.g. reactive agents vs. deliberative agentse.g. reactive agents vs. deliberative agents

Integration of Integration of at agent level at agent level e.g. layered architecturee.g. layered architecture

Interactions at Interactions at one level onlyone level only e.g. knowledge level e.g. knowledge level

A lot of scenarios can benefit from A lot of scenarios can benefit from hybrid solutionshybrid solutions

To address complete scenarios real infospheres To address complete scenarios real infospheres

have to have to overcomeovercome this this compartmentalizationcompartmentalization..

Toward complex information ecosystemsToward complex information ecosystems

A complex information ecosystem includes A complex information ecosystem includes chains chains and webs leveraging the variety of agentsand webs leveraging the variety of agents

Allows for a Allows for a pyramid of speciespyramid of species where where

each level brings some added valueeach level brings some added value

to the whole information chain;to the whole information chain;

Allows for a Allows for a large spectrumlarge spectrum of of

agent types addressing the largeagent types addressing the large

spectrum of information tasksspectrum of information tasks

required by scenarios of use;required by scenarios of use;

Allows Allows direct interactions and indirect chainsdirect interactions and indirect chains and and

webs of interactionswebs of interactions across the different levels. across the different levels.

UsersUsers

WebsWebs

Progression of the talkProgression of the talk

General positionGeneral position

Argue for complete ecosystems in infospheresArgue for complete ecosystems in infospheres

Technological stanceTechnological stance

Summarize interests of XML information landscapeSummarize interests of XML information landscape

Initial and current workInitial and current work

Examples of implementations and experimentsExamples of implementations and experiments

Technological stance: XML standardsTechnological stance: XML standards

XML: XML: shapingshaping distributed information landscapes. distributed information landscapes. Structure documents and data using a text format.Structure documents and data using a text format. Platform-independent, internationalization, Platform-independent, internationalization,

localization, validation, license-free, localization, validation, license-free, etc.etc. Distributed information that can be processed.Distributed information that can be processed.

RDF/S and OWL: encoding distributed RDF/S and OWL: encoding distributed semanticssemantics.. Annotate Annotate Web resourcesWeb resources with with properties/relationsproperties/relations.. Encode Encode ontologiesontologies for annotation and interactionsfor annotation and interactions..

XSLT+XPath: describe XSLT+XPath: describe modification modification of information.of information. Rule-based language for XML Rule-based language for XML tree transformationtree transformation.. Selecting, sorting, counting, variables, parameters, Selecting, sorting, counting, variables, parameters,

importingimporting other stylesheets, extensions, other stylesheets, extensions, etc.etc.

XSLT and agentsXSLT and agents

XML XML exchange format for structured data.exchange format for structured data.XSLT XSLT exchange format for data manipulation.exchange format for data manipulation.

XML: declarative language in agent com.XML: declarative language in agent com.XSLT: procedural language in agent com.XSLT: procedural language in agent com.

Two perspectives on XSLT in agent interactions Two perspectives on XSLT in agent interactions Dynamically customize generic information agent Dynamically customize generic information agent

roles at run-time roles at run-time // holonic approach// holonic approach.. Describe and propagate simple reactive XML Describe and propagate simple reactive XML

agents agents // ecosystem approach// ecosystem approach.. In both cases XSLT is used to propagate simple In both cases XSLT is used to propagate simple

XML manipulation behaviors.XML manipulation behaviors. Use Use standard protocolsstandard protocols for propagation for propagation

e.g. FIPA-Request e.g. FIPA-Request

XSLT AgentsXSLT Agents

Constructors provided by XSLT:Constructors provided by XSLT: Sensors = patterns of a template or the test Sensors = patterns of a template or the test

instructions both using the XPath expressions;instructions both using the XPath expressions; Effectors = the value-manipulating instructions;Effectors = the value-manipulating instructions; Reactions = recursive rules & branching instructionsReactions = recursive rules & branching instructions

Rule #1: respect the environmentRule #1: respect the environment

<xsl:stylesheet xmlns:xsl=" (...) ">

<xsl:template match="@rdf:about (...) "> <xsl:if test="not( (...) "> <xsl:value-of select="substring-before(., (...)

</xsl:template>

</xsl:stylesheet>

effectors

reflex

sensorsstimulus pattern

<xsl:template match="@*|node()"> <xsl:copy> <xsl:apply-templates select="@*|node()"/> </xsl:copy></xsl:template>

Progression of the talkProgression of the talk

General positionGeneral position

Argue for complete ecosystems in infospheresArgue for complete ecosystems in infospheres

Technological stanceTechnological stance

Summarize interests of XML information landscapeSummarize interests of XML information landscape

Initial and current workInitial and current work

Examples of implementations and experimentsExamples of implementations and experiments

Context of experimentationContext of experimentation

CoMMA (IST CoMMA (IST INRIA, ATOS-Origin, CSTB, LIRMM, Deutsche Telekom, Univ. ParmaINRIA, ATOS-Origin, CSTB, LIRMM, Deutsche Telekom, Univ. Parma )) Corporate memory as a Corporate memory as a corporate semanticcorporate semantic web web

= ontology + annotations of docs, org. & people= ontology + annotations of docs, org. & people Improve precision/recall, push, organize archivesImprove precision/recall, push, organize archives User Interface Controller; Profile Manager; Profile User Interface Controller; Profile Manager; Profile

Archivist; Archivist; Ontology Archivist; Corporate Model Ontology Archivist; Corporate Model Archivist; Archivist; Annotation Archivist; Annotation Mediator Annotation Archivist; Annotation Mediator

mymyCampus (DAML Campus (DAML CMU, DARPA, Boeing, HP, IBM, Symbol, Fujitsu, Amazon, CMU, DARPA, Boeing, HP, IBM, Symbol, Fujitsu, Amazon,

ISTIST )) Mobile accesses to Mobile accesses to context-aware servicescontext-aware services Open architecture: e-Wallets, User Interaction Open architecture: e-Wallets, User Interaction

Manager, Task-Specific Agents. Manager, Task-Specific Agents. Both cases: Semantic Web + Deliberative agentsBoth cases: Semantic Web + Deliberative agents

Customize behaviors: Web wrapperCustomize behaviors: Web wrapper Automate Automate extractionextraction of relevant pieces from Web of relevant pieces from Web

IntegrateIntegrate them to the organizational memory: them to the organizational memory: Sample page Sample page HTML HTML XHTML XHTML XHTML XHTML Example annotation Example annotation XSLT XSLT Create Annotation WrapperCreate Annotation Wrapper

(XSLT + Web sources)(XSLT + Web sources) Annotation MediatorAnnotation Mediator

(query solving & monitoring)(query solving & monitoring) XSLT for extraction taskXSLT for extraction task

Built-in templates = high level extraction functionsBuilt-in templates = high level extraction functions Composition, extension, propagationComposition, extension, propagation

Behavior of wrappers initially with generic task of Behavior of wrappers initially with generic task of extraction then customized at run-time extraction then customized at run-time // holonic// holonic

{URL}

1

1

1

1Interfacecontroller

Annotationmediator

Annotationwrapper &archivistA

rchive

Corporate WebOpen Web

CoMMA Wrapper generationCoMMA Wrapper generation

myCampus wrapped servicesmyCampus wrapped services

Customize behaviors: semantic gatewayCustomize behaviors: semantic gateway

Temporary Temporary extranetsextranets supporting a virtual supporting a virtual organization: organization: connection of semantic intrawebsconnection of semantic intrawebs

Generic gateway agentGeneric gateway agent

translating between translating between

ontologiesontologies of the different of the different

organizationsorganizations

Semi-automatic mapping construction (simple tf*icf)Semi-automatic mapping construction (simple tf*icf)

Set of XSLT templates to translate query/annotationSet of XSLT templates to translate query/annotation

Upload / customization of translation behavior Upload / customization of translation behavior Generic gateways for translation and securityGeneric gateways for translation and security

task of translation customized at run-time task of translation customized at run-time // holonic// holonic

1

*

Annotationmediator

Gatewaymediator

Corporate Web2Corporate Web1

Annotationmediator

1

*

Customize behaviors: dynamic interfacesCustomize behaviors: dynamic interfaces

Customizing and extending interfacesCustomizing and extending interfacesCoMMA:CoMMA:mymyYahoo-likeYahoo-likeontology-basedontology-basedqueryingquerying

Customize behaviors: dynamic interfacesCustomize behaviors: dynamic interfaces

Customizing and extending interfacesCustomizing and extending interfacesmymyCampus:Campus:Task-specificTask-specificagents’agents’interfacesinterfaces

Swarm propagation: maintenance Swarm propagation: maintenance Life-cycle of (distributed) knowledgeLife-cycle of (distributed) knowledge: update : update

annotations, maintain coherence, erase old ones…annotations, maintain coherence, erase old ones… Generate swarm agent e.g. update URI of resourceGenerate swarm agent e.g. update URI of resource

(…) <CoMMA:WebPage rdf:about="http://www.inria.fr/acacia/ "> <CoMMA:Title>Web page of ACACIA</CoMMA:Title> <CoMMA:CreatedBy> <CoMMA:Person rdf:about="http://www.inria.fr/~dieng/" /> </CoMMA:CreatedBy> </CoMMA:WebPage>(…)

(…)<CoMMA:WebPage rdf:about="http://www-sop.inria.fr/acacia/ "> <CoMMA:Title>Web page of ACACIA</CoMMA:Title> <CoMMA:CreatedBy> <CoMMA:Person rdf:about="http://www-sop.inria.fr/~dieng/" /> </CoMMA:CreatedBy> </CoMMA:WebPage>(…)

(…)<xsl:template match="@rdf:about[contains(.,'www.inria.fr')]"><xsl:attribute name="rdf:about"><xsl:value-of select="substring-before(.,'www.inria.fr')"/>www-sop.inria.fr<xsl:value-of select="substring-after(.,'www.inria.fr')"/></xsl:attribute></xsl:template>(…)

prop

agat

e

Swarm propagation: info. fermentation (I)Swarm propagation: info. fermentation (I)

Reactive agents and shallow processing, e.g.:Reactive agents and shallow processing, e.g.: Annotations added & archived in distributed basesAnnotations added & archived in distributed bases Reactive agents propagated to enrich annotationsReactive agents propagated to enrich annotations Query & push agents retrieve relevant annotationsQuery & push agents retrieve relevant annotations Interface agents display enriched resultsInterface agents display enriched results

Testing: PubMed from National Library of MedicineTesting: PubMed from National Library of Medicine 9981 annotations extracted by Annotation Wrapper9981 annotations extracted by Annotation Wrapper Behavior = cross-pollenizing/pollination; bee2bee Behavior = cross-pollenizing/pollination; bee2bee 1. XSLT script starts from an annotation1. XSLT script starts from an annotation & extracts its list of authors & extracts its list of authors

2. Propagate2. Propagate 3. For each other annotation visited3. For each other annotation visited leave pheromone if the visited annotation leave pheromone if the visited annotation shares authors with the initial annotation shares authors with the initial annotation

Swarm propagation: info. fermentation (II)Swarm propagation: info. fermentation (II)

Pheromone track left by pollination agent:Pheromone track left by pollination agent:

Over the 9981 annotations:Over the 9981 annotations: 7724 ‘sameAuthorAs’ links generated 7724 ‘sameAuthorAs’ links generated Linking 2728 reports together Linking 2728 reports together i.e.i.e. 27% of the base 27% of the base

<c:ResearchReport rdf:about="URL in visited annotation"> (…)<c:sameAuthorAs> <c:SameAuthorDoc c:nbSharedAuthors="nb shared authors" rdf:about="url initial document"/> </c:sameAuthorAs></c:ResearchReport>

used to ranks

one step in a track

1

10

100

1000

10000

1 10 100 1000# author overlap

# r

ep

ort

s

Biology191 authors

Swarm propagation: info. fermentation (III)Swarm propagation: info. fermentation (III)

UsersUsers

interfaceinterfacecontrollercontroller

webwebwrapperwrapper

semantic intrawebsemantic intraweb

annot.annot.mediatormediator

archivistarchivistontologistontologistpollinatorpollinator

ConclusionConclusion This is This is not aboutnot about expressiveness of expressiveness of XSLTXSLT, XPath..., XPath... BeyondBeyond organizational approaches purely organizational approaches purely

deliberative or reactivedeliberative or reactive.. InfosphereInfosphere = ecosystem with large diversity of = ecosystem with large diversity of

interactions between lots of different agent typesinteractions between lots of different agent typesto maintain & exploit information landscape.to maintain & exploit information landscape.

Two perspectives:Two perspectives: Intelligent agents tasks customized at runtime.Intelligent agents tasks customized at runtime.

Relying on standard protocols to exchange proc. k.Relying on standard protocols to exchange proc. k.// holonic approach// holonic approach

Intelligent agents farming swarm intelligenceIntelligent agents farming swarm intelligenceReactive agents encapsulate ad-hoc protocols.Reactive agents encapsulate ad-hoc protocols.// ecosystem approach// ecosystem approach

AcknowledgementsAcknowledgements

CoMMA - ACACIA LaroratoryCoMMA - ACACIA Laroratory INRIA Sophia Antipolis INRIA Sophia AntipolisIST Program, ATOS-Origin, CSELT/Telecom Italia, IST Program, ATOS-Origin, CSELT/Telecom Italia, CSTB, INRIA, LIRMM, T-Nova/Deutsche Telekom, CSTB, INRIA, LIRMM, T-Nova/Deutsche Telekom, University of Parma University of Parma

mymyCampus - Mobile Commerce LaboratoryCampus - Mobile Commerce Laboratory Carnegie Mellon University Carnegie Mellon UniversityDAML / DARPA, Carnegie Mellon University, DAML / DARPA, Carnegie Mellon University, Boeing, HP, IBM, Symbol, Fujitsu, Amazon,Boeing, HP, IBM, Symbol, Fujitsu, Amazon,IST (SWAP) IST (SWAP)