Title: Perspectives AAAI Symposium on Agent Mediated Knowledge Management March 26, 2003
1PerspectivesAAAI Symposium on Agent Mediated
Knowledge ManagementMarch 26, 2003
- Sidney Bailin
- Knowledge Evolution, Inc. (www.kevol.com)
- Walt Truszkowski
- NASA/Goddard Space Flight Center
2Problem Scenario
- Agents must cooperate in some task
- They disagree about what must be done
3A Possible Diagnosis and Treatment
- Ontology mismatch
- Agents mean different things by the same terms
- Agents view the world differently
- Solution ontology negotiation
- An incremental process for establishing shared
understanding - Cf. Bailin Truszkowski, Ontology negotiation
between intelligent information agents,
Knowledge Engineering Review, Volume 17, Issue 1,
March 2002.
4How do the agents know theres an ontology
mismatch?
5Simple Model of an Observer
Observer
Knowledge Base
Message Stream
Interpretation
6Making Sense of an Observation Iterative KB Update
p, q, r, s, t, u, v, w ...
Message
Update
a, b, c, d, e ?
Interpretation
Update Problem
7Examples of Observation
- Remote Sensing
- Auditory
- Name that tune
- What does this mean?
- Verbal
- Intelligence analysis
8Knowledge Representations
- Propositional
- Assertions
- Ontological
- Assertions about types
- Event sets
- E.g., Probabalistic models
- Attribute sets
- E.g., Neural net interpretations
- Lattice
- States of knowledge
9Propositional Representation
10Event Set Representation
11Attribute Set Representation
Observer
Schema - Type - Source - Capability -
Threat ...
Position 50? Velocity 600 km/hr 70? Size
3m Shape oblong ...
Type Missile Source Adversary Capability
Evasive Threat level High Auto-respond Yes ...
Interpretation
12Attribute Set Representation
13OntologicalRepresentation
14Integrated Model
Events in the initial interpretation process
are the possible attribute-values that can be
inferred about the observed phenomenon.
Ontology
Attributes
Categories
Type Missile Source Adversary Capability
Evasive Threat level High Auto-respond Yes ...
Message
Events in the classification process are the
possible categories to which the observed
phenomenon can be assigned.
Classification
Initial Interpretation
Uncertainty Method
15What is Perspective (1 of 2)
- Something that distinguishes one observation from
another - Must be a difference in input, interpretation
process, or knowledge base
16What is Perspective (2 of 2)
- Choice of input signal
- Signal type
- Wavelength
- State of observer, target of observation, and
environment - E.g., time, position, angle of observation
- As represented in the KB
- Interpretation function
- Choice of significant info in input signal
- Mapping of selected signal to KB ontology
- This is like humans perspective on an issue
17When Observations Conflict
- Problem with signal?
- Phenomenon noise
- Resolution method calibration
- Difference in state?
- Phenomenon state-dependent events
- Resolution method correlate state with events
- Difference in interpretation?
- Phenomenon ontology conflict
- Resolution method ontology negotiation
18Summary
- Weve developed a model of observation and
perspective - Agents can explicitly represent and reason about
perspective (theirs and others) - Agents can use perspective reasoning to determine
how to resolve a conflict - In particular, to determine if and when an
ontology conflict is occurring