Adaptive Knowledge Discovery Framework (AKDF)
Perakath Benjamin
Paul Koola
Satheesh Ramachandran
Madhav Erraguntla
Ronald Fernandes
Knowledge Based Systems, Inc.
1408 University Drive East
College Station, TX 77840
[email protected] Abstract distributed, multi-source, multi-format data; (2) data
mining mechanisms for extracting valuable decision
The Adaptive Knowledge Discovery Framework enabling knowledge; (3) information fusion based
(AKDF) combines machine learning and knowledge knowledge harmonization for addressing specific real-time
based techniques to facilitate real-time decision
planning decision support requirements; (4) knowledge
support. AKDF is innovative in its synthesis of data
mining techniques and ontology-driven methods in a capture and management to allow for knowledge
system that is both scalable and easily reconfigurable. acquisition and maintenance from human experts; and (5)
adaptive learning to improve knowledge discovery
performance over extended periods of time.
Key innovations include (1) a hybrid approach to
Introduction combine data mining techniques with ontology-driven
This paper describes the architecture and concept of methods for automated and rapid knowledge discovery, (2)
operation of an Adaptive Knowledge Discovery innovative self-adaptation mechanisms that, in response to
Framework (AKDF). The framework uses a hybrid rapidly changing application environments, ensure
approach that combines emerging machine learning automated revisions to knowledge bases and knowledge
techniques with traditional knowledge based techniques to deployment strategies, and (3) a scalable and re-
extract knowledge from distributed and disparate data configurable distributed agent architecture that enables
sources and to rapidly deploy this knowledge for robust, rapid and cost effective knowledge based development and
real-time decision support applications. The practical deployment.
utility of the framework is illustrated in the third section of
this paper using an example scenario of a military mission
planning decision support application.
AKDF Architecture Summary
The AKDF is a scalable and rapidly reconfigure-able
framework for agent based decision support. AKDF The AKDF functional architecture is shown in Figure 1.
capabilities include (1) automated pre-processing of
Knowledge
Knowledge Capture
Repository Tools
Agent Coordinator
Signals
Filtered
Cleaned Information
Data Data Information Knowledge
Pre- Extraction Discovery
Processing Agents Agents
Agents
Reconfiguration Actionable
Decision
Advise
Knowledge
Adaptive Performance
Distributed
Learning Evaluation
Data
Agents Monitored Agents
Sources
Performance
Figure 1. AKDF Functional Architecture
system that provides dynamic knowledge discovery that is
Data from distributed data sources (digital terrain
responsive to unpredictable variability in the target
databases, signal intelligence, new feeds, etc.) are cleaned,
operational environment.
repaired, and transformed through the data pre-processing
activity. Data and signal pre-processing methods are used
to clean, filter, and validate the raw data. The filtered data Example Application Scenario
is further “mined” using knowledge discovery methods The practical benefits of AKDF will be illustrated through
(clustering, classification, association discovery), a knowledge-based military mission planning application
producing decision-enabling information for planning. example. The AKDF application manages knowledge
Knowledge discovery model reconfigurations (neural net collected from distributed, multi-disciplinary sources
weight adjustments, genetic algorithm search strategy (including military and geospatial experts) and deploys this
refinements, simulated annealing temperature gradient knowledge for real time military planning decision support.
adjustments, etc.) are automatically performed based on The AKDF application has three main innovative
the continuous monitoring and assessment of performance capabilities: (1) Knowledge Elicitation, (2) Knowledge
metrics. The adaptive learning adjustments to model Application, and (3) Dynamic Knowledge Update (Figure
parameters are stored back in the knowledge base. The 2).
proposed solution is thus a closed loop feedback control
End Users
Multi-Source
Data Feeds
Dynamic
Knowledge Knowledge
Knowledge
Elicitation Application
Update
Knowledge Data
Bases Base
Figure 2. AKDF Application Architecture Solution Concept
The main functions of the AKDF application are
summarized in the following paragraphs.
1. Knowledge Elicitation includes knowledge capture,
representation, organization, encoding, storage, and
communication; formal knowledge representations
include Plan Templates, Ontologies, and Production
Rules. Knowledge disciplines include Military
Planning and Geospatial (including physical and
cultural geography).
2. Plan Assessment includes Conflict Detection, Plan
Performance Assessment, and GIS-based Analysis.
Assessment involves (a) identifying plan infeasibilities
and (b) evaluating the relative merit of multiple plans
(based on multiple criteria including Time and Risk).
3. Dynamic Knowledge Update allows end users to be
informed of emerging data and information that might
be relevant to the current decision support task
(mission planning, geospatial analysis, etc.).
