Docstoc

NCSA-ARL-08041999-demo

Document Sample
NCSA-ARL-08041999-demo Powered By Docstoc
					    Neural, Bayesian, and Evolutionary Systems
               for High-Performance
     Computational Knowledge Management:
                  Demonstrations

                        Wednesday, August 4, 1999

       William H. Hsu, Loretta Auvil, Tom Redman, Michael Welge
         Michael Bach, Peter L. Johnson, Mike Perry, Kristopher Wuollett


                       Automated Learning Group
            National Center for Supercomputing Applications
                       http://www.ncsa.uiuc.edu/STI/ALG

PET Program Year-End Review                                 University of Illinois at Urbana-Champaign
                  Overview: Tools for Dealing
                  with Multisensor T&E Data
 • Short-Term Objectives: Building a Data Model
     – Progress to date: data channel typing for ontology
     – Current work: CGI form for data channel grouping, selection
     – Future work: integrity-checking, data preparation modules
 • Longer-Term Objectives
     – Multimodal Sensor Integration: multiple models in data fusion itinerary
     – Relevance Determination: genetic algorithm wrapper (current work)
     – Causal (Explanatory) Models: Bayesian network based on ontology
 • Test Bed: Super ADOCS Data Format (SDF)
     – 1719-channel asynchronous data bus (General Dynamics)
     – Experiment/Data Design
        – Typing: interactive tool for constructing data model
        – Specification of prediction target based on caution/warning channels
        – Interactive specification tool for learning architectures, algorithms
        – Target end users: test/instrumentation report designers, implementors
     – Analytical Applications: Decision Support
PET Program Year-End Review                                    University of Illinois at Urbana-Champaign
           Super ADOCS Data Format (SDF)
        Data Conversion and Selection Interface




 • CGI (Perl-based) form: Apache, MS Internet Explorer 5
PET Program Year-End Review                        University of Illinois at Urbana-Champaign
                    An Ontology for T&E Data

 •   Application Testbed
      – Aberdeen Test Center: M1 Abrams main battle tank (SEP data, SDF)
      – Generic Data Model (Facility for Experiment Specification)
 •   T&E Information Systems: Common Characteristics
      – Large-Scale Data Model
         • Objective: develop system capable of reducing model complexity
         • Methodology: build a relational (taxonomic, definitional) model of data
     – Data Integrity Requirements
         • Interactive form-based specification of test objective
         • Specification of error metrics, visualization criteria
     – Multimodality
         • Selection of relevant data channels
         • Interactive, support for automation
     – Data Reduction Requirements
         • Non-uniform downsampling - requires database of engineering units
         • Irrelevant data channels - requires type hierarchy


PET Program Year-End Review                                         University of Illinois at Urbana-Champaign
                             SDF Ontology:
                           Data Channel Types




                                                                              Spatial/GPS/
Caution/Warning     Profilometer      Fuel Systems    Timing
                                                                               Navigation
                  Data Bus/Control/
   Hydraulics                          Ballistics    Electrical                   Unused
                    Diagnostics


PET Program Year-End Review                                    University of Illinois at Urbana-Champaign
                Intranet Operating Environment

 •   Database Access
      – SDF import, flat file export
      – Internal data model: interaction with learning modules
      – Future development: SQL/Oracle 8 (JDBC) interface
 •   Deployment
      – CGI, JavaScript stand-alone applications
      – Management of modules, data flow through forms
 •   Presentation: Web-Based Interface
      – Simple, HTML-based invocation system
         • Common Gateway Interface (CGI) and Perl
         • Alternative implementation: servlets (http://www.javasoft.com)
      – Configuration of data model (file generation)
      – Management of experiments
         • Construction of models
         • Specification of learning systems (model architecture, training algorithm)
 •   Messaging Systems (Deployment  Presentation)

PET Program Year-End Review                                      University of Illinois at Urbana-Champaign
             Super ADOCS Data Format (SDF)
               Experiment Design Interface




 • D2K Genetic “Wrapper” for Data Channel Selection
PET Program Year-End Review                      University of Illinois at Urbana-Champaign
         Time Series Analysis and Visualization:
                   System Prototype




 • Visible Decisions Inc. (VDI) In3D
PET Program Year-End Review             University of Illinois at Urbana-Champaign
                        Summary and Conclusion

•   Model Identification
                                                            60
     – Queries: test/instrumentation reports
     – Specification of data model                          50


     – Grouping of data channels by type                    40




                                               Effort (%)
•   Prediction Objective Identification                     30


     – Specification of test objective                      20


     – Identification of metrics                            10


•   Reduction                                               0
                                                                   Objective     Data Preparation    Machine          Analysis &
     – Refinement of data model                                  Determination                       Learning        Assimilation

     – Selection of relevant data channels (given prediction objective)
•   Synthesis: New Data Channels
•   Integration: Multiple Time Series Data Sources

         Environment         Learning       Knowledge                                 Time Series
         (Data Model)        Element          Base                                 Analysis/Prediction



PET Program Year-End Review                                                                University of Illinois at Urbana-Champaign

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:1
posted:2/14/2012
language:English
pages:9