Building Intelligent
Systems
tell
Marie desJardins
Tim Finin
Anupam Joshi
Yun Peng
Yelena Yesha
register
tell
register
September 2004
1
Overview



Faculty: Finin, Yesha, Joshi, Peng, desJardins
Students: ~12 PhD, ~12 MS, ~5 undergrad
Labs:



Ebiquity: agents and the semantic web for open,
heterogeneous, distributed systems
MAPLE: MultiAgent Planning and LEarning
Funding


Current: DARPA (DAML), NSF (three ITRs, …),
Intelligence community, NASA, NIST, Industry, …
Recent: DARPA (CoABS, GENOA II)
UMBC
an Honors University in Maryland
2
ebiquity Group
KR
multiagent
systems
user
modeling
semantic
web services
web
AI Intelligent DB
data
machine learning
Information management
Systems
service discovery
and composition
wireless
Networking
mobility
& Systems
pervasive
computing
UMBC
an Honors University in Maryland
policy
Security
privacy
assurance
trust
3
ebiquity Group
KR
user
modeling
semantic
web
Building intelligent
AI Intelligent DB
systems
in
open,
Information
Systems
heterogeneous,
dynamic,
Networking
Security
& Systems
distributed
environments
multiagent
systems
machine learning
web services
data
management
service discovery and
composition
wireless
mobility
pervasive
computing
UMBC
an Honors University in Maryland
policy
assurance
privacy
trust
4
Some Relevant Project Areas
(1)
(2)
(3)
(4)
(5)
(6)
Agents and the semantic web
Policies for controlling autonomous agents
Context aware pervasive computing
TIVO for mobile computing
Trust in information systems
Large scale semantic web systems
UMBC
an Honors University in Maryland
5
(1) The Celebrity Couple
Semantic
Web
Software
Agents
In 2002, Geek Gossip gushed “The semantic web will
provide content for internet agents, and agents will make the
semantic web “come alive”. Looks like a match made in
Heaven!”
UMBC
an Honors University in Maryland
6
(1) Trading Agents

We’ve built an agent-based environment inspired by
TAC, the Trading Agent Competition




TAC is a forum for dynamic trading agent research with
games run in the last five years
TAC Classic involves a travel procurement, with agents
buying and selling goods for clients and scored on the cost
and clients’ preferences for trips assembled.
TAC is organized around a central auction server
Our goal was to open up the system, allowing peer-topeer communication among agents as well various kinds
of mediator, auction, discovery, service provider agents
… and to see how well the semantic web works as the
common knowledge infrastructure.
UMBC
an Honors University in Maryland
7
TAGA: Travel Agent Game inOwlAgentcities
for
Owl for
protocol
contract
Features
Technologies
Ontologies
descriptionhttp://taga.umbc.edu/ontologies/
enforcement
Open Market Framework
FIPA (JADE, April Agent Platform)
Motivation
Market dynamics
Auction theory (TAC)
Semantic web
Agent collaboration (FIPA &
Agentcities)
Owl for
modeling
trust
Auction Services
OWL message content
OWL Ontologies
Global Agent Community
Owl for
publishing
communicative
acts
travel.owl – travel concepts
fipaowl.owl – FIPA content lang.
auction.owl – auction services
tagaql.owl – query language
Semantic Web (RDF, OWL)
Web (SOAP,WSDL,DAML-S)
Internet (Java Web Start )
Owl for
representation
and reasoning
Owl for
negotiation
Report Direct Buy Transactions
Report Contract
Report Auction Transactions
Market Oversight
Agent
Bulletin Board
Agent
Customer
Agent
Report Travel Package
Auction Service
Agent
Proposal
Direct Buy
Web Service
Owl as a
Agents
content
FIPA platform
infrastructure services, including directory facilitators enhanced to use OWL-S for service discovery
language
Owl for
Owl for
authorization
service
policies
descriptions
Travel Agents
UMBC
an Honors University in Maryland
http://taga.umbc.edu/
8
What we learned

OWL is a good KR language for a reasonably
sophisticated MAS


OWL made it easy to mix content from different
ontologies unambiguously


Supporting partial understanding & extensibility
The use of OWL supported web integration


Integrates well with FIPA standards
Using information published on web pages and
integrating with web services via WSDL and SOAP
OWL has limitations: no rules, no default
reasoning, graph semantics, …

Some of which are being addressed
UMBC
an Honors University in Maryland
9
(2) It’s policies all the way down
1 A robot may not injure a
human being, or, through
inaction, allow a human
being to come to harm.
2 A robot must obey the
orders given it by human
beings except where such
orders would conflict with
the First Law.
3 A robot must protect its
own existence as long as
such protection does not
conflict with the First or
Second Law.
- Handbook of Robotics, 56th
Edition, 2058 A.D.
UMBC
an Honors University in Maryland
10
(2) It’s policies all the way down
1 A robot may not injure
a human being, or,
through inaction, allow
a human being to come
 Unlike traditional “hard coded” rules like
to harm.
DB access control & OS file permissions
2 A robot must obey the
 Autonomous agents need policies as
orders given it by human beings except
“norms of behavior” to be followed to
where such orders
be good citizens
would conflict with the
First Law.
 So, it’s natural to worry about …
 How agents governed by multiple policies 3 A robot must protect its
own existence as long
can resolve conflicts among them
as such protection does
 How to deal with failure to follow policies
not conflict with the
First or Second Law.
– sanctions, reputation, etc.
- Handbook of Robotics,
 Whether policy engineering will be any
56th Edition, 2058 A.D.
easier than software engineering
 In
Asimov’s world, the robots didn’t
always strictly follow their policies
UMBC
an Honors University in Maryland
11
Our Approach
 Policies

are useful at virtually all levels
OS, networking, data management, applications
 Declarative
policies guide the behavior of
entities in open, distributed environments
 Positive
& negative authorizations & obligations
 Focused on domain actions
 Policies are based on attributes of the action
(and its actor and target) and the general
context – not just on their identity of the actor
UMBC
an Honors University in Maryland
12
Rei Policy Language
 Developed
several versions of Rei, a policy
specification language, encoded in (1)
Prolog, (2) RDFS, (3) OWL
 Used to model different kinds of policies
 Authorization
for services
 Privacy in pervasive computing and the web
 Conversations between agents
 Team formation, collaboration & maintenance
 The
OWL grounding enables policies that
reason over SW descriptions of actions,
agents, targets and context
UMBC
an Honors University in Maryland
13
Rei Policy Language






Rei is a declarative policy language for describing policies
over actions
 Reasons over domain dependent information
Currently represented in OWL + logical variables
Based on deontic concepts
 Permission, Prohibition, Obligation, Dispensation
Models speech acts
 Delegation, Revocation, Request, Cancel
Meta policies
 Priority, modality preference
Policy engineering tools
 Reasoner, IDE for Rei policies in Eclipse
UMBC
an Honors University in Maryland
14
Applications – past, present & future
 Coordinating
access in supply chain
1999
 Authorization
policies in a pervasive
2002
management system
computing environment
 Policies
for team formation, collaboration,
information flow in multi-agent systems
 Security in semantic web services
 Privacy and trust on the Internet
 Privacy in pervasive computing
2003
…
2004
…
environments
UMBC
an Honors University in Maryland
15
What we learned




Declarative policies can be used to model
security, trust and privacy constraints
Reasonably expressive policy languages can be
encoded on OWL
This enables policies to depend on attributes
and context information available on the
semantic web
Policies are applicable at almost every level of
the stack, from systems and networking to
multiagent applications.
UMBC
an Honors University in Maryland
16
(3) A Love Triangle?
Semantic
Web
Software
Agents
Pervasive
Computing
Even matches made in Heaven don’t always work out as
planned.
UMBC
an Honors University in Maryland
17
(3) Pervasive Computing
“The most profound technologies are those
that disappear. They weave themselves into
the fabric of everyday life until they are
indistinguishable from it ” – Mark Weiser
Think: writing, central heating, electric
lighting, water services, …
Not: taking your laptop to the beach, or
immersing yourself into a virtual reality
UMBC
an Honors University in Maryland
18
UMBC
an Honors University in Maryland
19
UMBC
an Honors University in Maryland
20
Pervasive environments for the Military
UMBC
an Honors University in Maryland
21
This is a challenging environment
 While
devices are getting smaller, cheaper and
more powerful, they still have severe limitations.
Battery, memory, computation, connection, bandwidth
 Each as limited sensors and perspective

 The
environment is inherently dynamic with
serendipitous connections and unknown entities

This makes security and trust important
 MANETS
(mobile ad hoc networks) underlie
pervasive infrastructures like Bluetooth

It’s autonomous agents all the way down
 Security

and privacy is a special concern
Warfighters and agents must control how information
about them is collected and used
UMBC
an Honors University in Maryland
22
Representing and reasoning about context
CoBrA: a broker centric agent architecture
for supporting pervasive context-aware
systems
 Using
SW ontologies for context modeling
and reasoning about devices, space, time,
people, preferences, meetings, etc.
 Using logical inference to interpret context
and to detect and resolve inconsistent
knowledge
 Allowing users to define policies controlling
how information about them is used and
shared
UMBC
an Honors University in Maryland
23
A Bird’s Eye View of CoBrA
UMBC
an Honors University in Maryland
24
SOUPA Ontology provides common vocabulary
UMBC
an Honors University in Maryland
25
A Simple Spatial Model of UMBC
UMBC
an Honors University in Maryland
26
Where’s Harry?
UMBC
an Honors University in Maryland
27
Detecting Inconsistencies
UMBC
an Honors University in Maryland
28
Powerful AI tools
 DL
reasoning over privacy policies
 Abductive reasoning

Cobra uses a simple form of abductive reasoning to
keep track of assumptions underlying it’s conclusions
e.g.: A person’s at X if his cell phone’s at X assuming it’s not lost

This allows Cobra to find consistent and plausible
explanations for the data it sees.
 Argumentation

Agents can question or challenge each others beliefs,
engaging in simple argumentation dialogs
e.g.: giving the assumptions and provenance underlying a belief

An agent may choose not to adopt another’s belief
based on the provenance of the data
UMBC
an Honors University in Maryland
29
Privacy Protection in CoBrA
Users define policies to permit or prohibit
the sharing of their information
 Policies are provided by personal agents
or published on web pages
 and use the SOUPA ontologies as well
as other SW assertions (e.g., FOAF,
schedules)
 The context broker follows user defined
policies when sharing information, unless
contravened by higher policies
UMBC

an Honors University in Maryland
30
The SOUPA Policy Ontology
UMBC
an Honors University in Maryland
31
Policy Reasoning Use Case
 The
speaker doesn’t want others to know the
specific room that he’s in, but is willing for others
to know he’s on campus
 He defines the following privacy policy

Share my location with a granularity >= “State”
 The
broker
isLocated(US) => Yes!
 isLocated(Maryland) => Yes!
 isLocated(UMBC) => Uncertain..
 isLocated(ITE-RM210) => Uncertain..

UMBC
an Honors University in Maryland
32
What we learned




FIPA and OWL were good for integrating
disparate components, even on cell phones!
OWL made it easy to mix content from different
ontologies unambiguously
OWL made it easy to take advantage of
information published in XML on the web
 e.g., foaf information, privacy policy
Powerful AI tools add value

DL reasoning, abduction, argumentation
UMBC
an Honors University in Maryland
33
(4) TIVO for Mobile Computing
A mobile computing vision and a problem
 Devices
“broadcast” information and service
descriptions via short-range RF (802.11, Bluetooth,
UWB, etc.)
 As people and their devices move, they can access
this data, but only while it’s in range

The data may be out of range when it’s needed
 Devices
must anticipate their information need so
they can cache data when it’s available
Based on user model, preferences, schedule, context,
trust, …
 Compute a dynamic utility function to create a
“semantic” cache replacement algorithm

UMBC
an Honors University in Maryland
34
MoGATU’s distributed belief model
MoGATU is a data management module for MANETs
 Devices send queries to peers


Ask its vicinity for reputation of untrusted peers that responded -trust a device if trusted before or if enough trusted peers trust it
Use answers from (recommended to be) trusted
peers to determine answer
 Update reputation/trust level for all responding
devices




Trust level increases for devices giving what becomes final answer
Trust level decreases for devices giving “wrong” answer
Each devices builds a ring of trust…
UMBC
an Honors University in Maryland
35
B: I know
where Bob is.
C: I know
where Bob is.
A: Where is Bob?
D: I know
where Bob is.
UMBC
an Honors University in Maryland
36
A: B, where is Bob? A: C, where is Bob? A: D, where is Bob?
UMBC
an Honors University in Maryland
37
B: A, Bob is home.
C:
A, Bob is at work.
D:
A, Bob is home.
UMBC
an Honors University in Maryland
38
A:
B: Bob at home,
C: Bob at work,
D: Bob at home
A: I have enough
trust in D. What
about B and C?
UMBC
an Honors University in Maryland
39
B: I am not sure.
C: I always do.
F: I do.
E: I don’t.
A: Do you trust C?
D: I don’t.
UMBC
an Honors University in Maryland
A:
I don’t care what C says.
I don’t know enough about B,
but I trust D, E, and F. Together,
they don’t trust C, so won’t I.
40
B: I do.
C: I never do.
F: I am not sure.
E: I do.
A: Do you trust B?
D: I am not sure.
UMBC
an Honors University in Maryland
A:
I don’t care what B says.
I don’t trust C,
but I trust D, E, and F. Together,
they trust B a little, so will I.
41
A: I trust B and D,
both say Bob is
home…
A:
Bob is home!
UMBC
an Honors University in Maryland
A:
Increase trust in B.
A:
Decrease trust in C.
A:
Increase trust in D.
42
What we learned



OWL was a good language for capturing user
profiles and the simple BDI models we needed
Any of several simple trust models increase the
accuracy of information
 Designing a good trust model depends on the
MANET assumptions
 As well as the level of cooperation and
honesty
Trading reputation information boosts the
performance of the algorithms
UMBC
an Honors University in Maryland
43
(4) SEMDIS
Knowledge Discovery
in the Semantic Web
Objective
Approach
Design, prototype and evaluate a system supporting
the discovery, indexing and querying of complex
semantic relationships in the Semantic Web. The
system maintains and utilizes trust and provenance
Knowledge representation systems reason over semantic web content discovered on the web which is reduced to triples that can be efficiently stored and processed in relational databases. Trust models and
heuristics guide the formation of conclusions
information to enhance the relationship discovery.
xsd:real [0,1]
confidence
AssociationConnective
Association
Justification
foaf:Document
rdf:Resource
foaf:page
Trust
Belief
Reference
Broader impacts
Techniques and prototypes developed can be applied
to a range of problems, including discovering new
connections and relations in scientific information
and homeland security.
A “web of belief” model
and associated ontology is
used to represent, integrate,
and evaluate conclusions
drawn from the large
volume of heterogeneous
assertions found in the data.
connective
DocumentRelation
NSF award ITR-IIS-0325464
U. Georgia, Sheth, Arpinar, Kochut, Miller
NSF award ITR-IIS-0325172
UMBC, Joshi, Yesha, Finin
foaf:Agent
selects
contains
rdf:Statement
FOAF Network
Y. Yesha
island
source
Kagal
source
J. Golbeck
knows
L. Ding
H. Chen
J. Hendler
knows
P. Kolari
knows
F. Perich
T. Finin
A. Joshi
Golbeck’s
Trust Network
hub
sink
mapTo
Ding
Y. Peng
1
Kagal
Finin
28
6
A. Sheth
A. Joshi
1
Chen
SWETO is large ontology covering several test-bed domains. It is pop-ulated
with 800K instances and 1.M relations extracted from heterogeneous Web
sources. SWETO was developed using Semagix Freedom system.
5
An experimental algorithm has been developed to integrate and
rank discovered relationships.
M. P. Singh
Perich
DBLP Network
http://lsdis.cs.uga.edu/Projects/SemDis
June 2004
http://semdis.umbc.edu/
44
Need a slide or two here….
UMBC
an Honors University in Maryland
45
Approach
We are building prototype tools and applications that
demonstrate how semantic web technology supports information discovery, integration and sharing in scientific communities. The National Biological Information Infrastructure
(NBII) and Invasive Species Forecasting System (ISFS) provide requirements and serve as testbeds for our prototypes.
(5) SPIRE
Invasive species do more economic damage to the U.S. every year that all other natural disasters combined. Above: plants, animals, and a virus.
Semantic Prototypes in Research Ecoinfomatics
Spire is a distributed, interdisciplinary research project exploring how semantic web technology supports information discovery, integration, and sharing in scientific communities. We are building prototype tools and applications for inclusion in the
National Biological Information Infrastructure (NBII), with a focus on the early detection and warning of invasive species.
Meal of a Meal (after Friend of a Friend). We know
Fish 1 eats Plant 1. We then infer that Fish 1 may
also eat the taxonomic siblings of Plant 1: Plants 2
and 3. Similarly, we infer that the taxonomic
siblings of Fish 1 - Fishes 2 and 3 - may eat Plant 1.
The RMBL team expresses food webs in OWL using an ontology for ecological interaction they have constructed in coordination with other ecologists. The OWL model drives the simulation and visualization.
Significant Results
•
•
•
•
•
•
SWOOGLE - a search engine for the semantic web.
MoaM (Meal of a Meal) - Given a species list, infer a food web.
Photostuff - annotate regions of a picture with OWL.
SWOOP - the first ontology editor written specifically for OWL.
Ontologies for ecological interaction, and observation data.
Food web visualization and analysis tools that are driven by OWL
ontologies and instance data.
• CRISIS CAT - an RDF based catalog of Invasive Species
resources in California.
• Coordination with USGS, NASA, EPA, GBIF, and the
Intergovernmental, Interagency Cooperation on Ecoinformatics.
Swoogle is a crawler based search and retrieval system for semantic web documents (SWDs) in RDF and OWL. It discovers SWDs and computes their
metadata and relations, and stores them in an IR system. Users can search for
ontologies or instance data, and hits are ranked according to our Ontology
Rank algorithm.
Broader Impacts
•
Enable knowledge from one community to be effectively
used by another.
• Harness the power of the citizen scientist. (The majority of
invasives are discovered by amateurs.)
• Integrate research and education in the classroom.
Coming Soon
• ELVIS – an end to end application that starts with a location
and produces a model of its food web.
• The Pond Project - a junior high school classroom activity
to monitor the health of local ecosystems.
• Enhanced tools.
An ontology (found via Swoogle) is loaded
into Photostuff to mark up regions of a
field photograph.
The NBII California Information Node (CAIN), maintained by
UC Davis, is a jumping off point to broader NBII deployment.
UMBC
AN HONORS UNIVERSITY IN MARYLAND
Spatial distribution of exotic plants at the Cerro Grande
fire site. The statistical techniques used to generate
these maps do not take trophic data as input. Yet.
Research Team
UMBC ebiquity (Finin)
UMBC GEST Center (Sachs)
UMD MINDSWAP (Hendler)
UC Davis ICE (Quinn)
RMBL PEaCE (Martinez)
NASA GSFC (Schnase)
46
Research support was provided by NSF, award NSF-ITR-IIS-0326460, PI Tim Finin, UMBC.
CGI scripts
SWOs
Video
files
HTML
documents
SWIs
Web
services
Swoogle is a crawler based search & retrieval
system for semantic web documents (SWDs)
in RDF, Owl and DAML. It discovers SWDs
and computes their metadata and relations,
and stores them in an IR system.
SWD Properties
Language and level; encoding, number of
triples, defined classes, defined properties, &
defined individuals; type (SWO, SWI); form
(RSS, FOAF, P3P, …); rank; weight;
annotations; …
Web
interface
Ontology
Analyzer
Jena
Ontology
Ontology
Ontology
Agents
Ontology
Agents
Agents
Agents
APIs
Images
Agent
services
Apache/
Tomcat
php, myAdmin
mySQL
Jena
IR
engine
SIRE
SWD = SWO + SWI
Focused
Crawler
SWD
crawler
DB
Audio
files
Ontology
discovery
We
b
The web, like Gaul, is divided into three parts:
the regular web (e.g. HTML), Seman- tic
Web Ontologies (SWOs), and Semantic Web
Instance files (SWIs)
Ontology
Google
discovery
cached
files
SWD Relations
Binary: R(D1,D2)
• IM: D1 owl:imports D2
• IMstar: transitive closure of IM
• EX: D1 extends D2 by defining classes or
properties subsumed by D2’s
• PV: owl:priorVersion & subproperties
• TM: D1 uses terms from D2
• IN: D1 uses individual defined in D2
• MAP: D1 maps some of its terms to D2’s
• SIM: D1 & D2 are similar
• EQ: D1 & D2 are identical
• EQV: D1 & D2 have the same triples
Ternary: R(D1,D2,D3)
• MP3: D1 maps a term from D2 to D3 using
owl:sameClass, etc.
Swoogle uses two kinds of crawlers to discover semantic web documents and
several analysis agents to compute metadata and relations among documents
and ontologies. Metadata is stored in a relational DBMS.
SWD Rank
A SWD’s rank is a function of its type
(SWO/SWI) and the rank and types of the
documents to which it’s related.
http://swoogle.umbc.edu/
Swoogle has metadata on classes, properties
and individuals from ~240,000 SWDs
SWD IR Engine
Swoogle puts documents into a character ngram based IR engine to compute document
similarity and do retrieval from queries
Contributors include Tim Finin, Anupam Joshi, Yun Peng, R. Scott Cost, Jim Mayfield, Joel Sachs, Pavan Reddivari, Vishal Doshi,
Rong Pan, Li Ding, and Drew Ogle. Partial research support was provided by DARPA contract F30602-00-0591 and by NSF by
awards NSF-ITR-IIS-0326460 and NSF-ITR-IDM-0219649. 20 May 2004.
47
UMBC
an Honors University in Maryland
48
Uncertainty in Ontology Engineering:
A Bayesian Perspective
Yun Peng, Zhongli Ding, Rong Pan
Department of Computer Science and
Electrical engineering
University of Maryland Baltimore County
ypeng@umbc.edu
UMBC
an Honors University in Maryland
49
Motivations
• Uncertainty in ontology engineering
– In representing/modeling the domain
• Degree of inclusion
• Overlap rather than inclusion
• Uncertain information often available
– In reasoning
• How close a description D is to its most specific subsumer?
• Noisy data: leads to over generalization in subsumptions
• Uncertain input:
– In mapping concepts from one ontology to another
• Similarity between concepts may not be described logically
• Mappings are hardly 1-to-1
• Uncertainty becomes more prevalent in web environment
– One ontology may import other ontologies
– Competing ontologies for the same or overlapped domain
UMBC
an Honors University in Maryland
50
Overview of The Approach
onto2
onto1
Probabilistic
ontological
information
P-onto2
P-onto1
Probabilistic
annotation
BN1
Probabilistic
ontological
information
BN2
OWL-BN
translation
– OWL-BN translation
concept
mapping
• By a set of translation rules and
procedures
• Maintain OWL semantics
• Ontology reasoning by probabilistic
inference in BN
– Ontology mapping
• A parsimonious set of links
• Capture similarity between concepts
by joint distribution
• Mapping as evidential reasoning
UMBC
an Honors University in Maryland
51
OWL-BN Translation
• Translated BN will preserve
– Semantics of the original ontology
– Encoded probability distributions among relevant variables
• Encoding probabilities in OWL ontologies
– Define new OWL classes for prior and conditional
probabilities
• Structural translation: a set of rules
– Class hierarchy: set theoretic approach
• Each OWL class translated to a BN node
• Arcs from super to subclass nodes
– Logical relations (equivalence, disjoint, union, intersection...)
• Control nodes whose CPT realize logical relations
– Properties (ongoing)
UMBC
an Honors University in Maryland
52
Structural Translation
• Set theoretic approach
– Each OWL class is considered a set of objects/instances
– Each class is defined as a node in BN
– An arc in BN goes from a superset to a subset
– Consistent with OWL semantics
<owl:Class rdf:ID=“Human">
<rdfs:subclassOf rdf:resource="#Animal">
<rdfs:subclassOf rdf:resource="#Biped">
</owl:Class>
RDF Triples:
(Human rdf:type owl:Class)
(Human rdfs:subClassOf Animal)
(Human rdfs:subClassOf Biped)
Translated to BN
UMBC
an Honors University in Maryland
53
Structural Translation
• Logical relations
– Some can be encoded by CPT (e.g.. Man = Human∩Male)
– Others can be realized by
adding control nodes
Man  Human
Woman  Human
Human = Man  Woman
Man ∩ Woman = 
auxiliary node: Human_1
Control nodes: Disjoint, Equivalent
UMBC
an Honors University in Maryland
54
Constructing CPT
• Imported Probability information is not in the form of CPT
• Assign initial CPT to the translated structure by some default
rules
• Iteratively modify CPT to fit imported probabilities while
setting control nodes to true.
– IPFP (Iterative Proportional Fitting Procedure)
To find Q(x) that fit Q(E1), … Q(Ek) to the given P(x)
• Q0(x) = P(x); then repeat Qi(x) = Qi-1(x) Q(Ej)/ Qi-1(Ej) until
converging
• Q (x) is an I-projection of P (x) on Q(E1), … Q(Ek) (minimizing
Kullback-Leibler distance to P)
– Modified IPFP for BN
UMBC
an Honors University in Maryland
55
Example
UMBC
an Honors University in Maryland
56
Ontology Mapping
• Formalize the notion of mapping
• Mapping involving multiple concepts
• Reasoning under ontology mapping
UMBC
an Honors University in Maryland
57
Formalize The Notion of Mapping
• Simplest case: Map concept E1 in Onto1 to E2 in Onto2
– How similar between E1 and E2
– How to impose belief (distribution) of E1 to Onto2
• Cannot do it by simple Bayesian conditioning
P(x| E1) = ΣE2 P(x| E2)P(E2 | E1) similarity(E1, E2)
– Onto1 and Onto2 have different probability space (Q and P)
• Q(E1) ≠ P(E1)
• New distribution, given E1 in Onto1: P*(x) ≠ΣP (x|E1)P(E1)
– similarity(E1, E2) also needs to be formalized
UMBC
an Honors University in Maryland
58
Formalize The Notion of Mapping
• Jeffrey’s rule
– Conditioning cross prob. spaces
– P*(x) =ΣP (x|E1)Q(E1)
– P* is an I-projection of P (x) on Q(E1) (minimizing KullbackLeibler distance to P)
– Update P to P* by applying Q(E1) as soft evidence in BN
• similarity(E1, E2)
– Represented as joint prob. R(E1, E2) in another space R
– Can be obtained by learning or from user
• Define
map(E1, E2) = <E1, E2, BN1, BN2, R(E1, E2)>
UMBC
an Honors University in Maryland
59
Reasoning With map(E1, E2)
BN1
Q
P
E2
E1
Applying Q(E1) as
soft evidence to
update R to R* by
Jeffrey’s rule
BN2
R
E1
E2
Applying R*(E2)
as soft evidence to
update P to P* by
Jeffrey’s rule
Using similarity(E1, E2):
R*(E2)
= R*(E1, E2)/R*(E1)
UMBC
an Honors University in Maryland
60
Mapping Reduction
• Multiple mappings
– One node in BN1 can map to all nodes in BN2
– Most mappings with little similarity
– Which of them can be removed without affecting the overall
• Similarity measure:
– Jaccard-coefficient: sim(E1, E2) = P(E1  E2)/R(E1  E2)
– A generalization of subsumption
– Remove those mappings with very small sim value
• Question: can we further remove other mappings
– Utilizing knowledge in BN
UMBC
an Honors University in Maryland
61
Current focuses
• Current focuses:
– OWL-BN translation: properties
– Algorithms for ontology reasoning as probabilistic inference
over translated BN
– Ontology mapping: mapping reduction
• Prototyping and experiments
• Issues
– Complexity
– How to get these probabilities
UMBC
an Honors University in Maryland
62
UMBC
an Honors University in Maryland
63
Marie’s slides go here
UMBC
an Honors University in Maryland
64
backup
UMBC
an Honors University in Maryland
65