Ant Colony Optimisation for Fuzzy Rule Induction

Document Sample
Ant Colony Optimisation for Fuzzy Rule Induction Powered By Docstoc
					      Encouraging Complementary
          Fuzzy Rules within
        Iterative Rule Learning

  Michelle Galea                     Qiang Shen
 School of Informatics      Department of Computer Science
University of Edinburgh           University of Wales
    Edinburgh, UK                  Aberystwyth, UK

                      Vishal Singh
               Larson & Toubro, EmSys Ltd.
                     Bangalore, India
                Motivation

1.   Gain deeper understanding of IRL
     strategy for fuzzy rule base induction

2.   Test ACO as rule discovery mechanism
     within IRL
IRL – Iterative Rule Learning

                         SPBA1
           adjustments           best rule
                                             Rule Base
                                              Rule 1
Training                         best rule    Rule 2
                         SPBA2
   Set                                          .
           adjustments                          .
                           .                  Rule k
                           .
                           .
                                 best rule


                         SPBAk
  Ant Colony Optimisation –
         The Basics
Constructionist, iterative algorithm:
     Problem representation
     Probabilistic transition rule
     Local heuristic
     Constraint satisfaction method
     Fitness function
     Pheromone updating strategy
         ACO for Fuzzy Rule Induction
                                             ACO 1

        Iteration 1                         Iteration 2            ..........                 Iteration m

Rule                             Rule                                              Rule
 1 .1                             2 .1                                              m.1
        Rule        Rule                    Rule           Rule                               Rule     Rule
         1.2         1. n                    2.2            2. n                               m.2      m. n



                                 best rule itn. 2
               best rule itn.1                                             best rule itn. m

                                         Rule 1.2
                                         Rule. 2.5
                                             .
                                         Rule m.3

                                                                   Rule base
                                                     best rule      Rule 1
FRANTIC Rule Construction…
                   OUTLOOK
HUMIDITY


                              Cloudy
 Not_H             Rain
           Humid
                          Sunny




  Cool
           Hot

   Mild


TEMPERATURE        Not_W
                                  Wind


                           WIND
FRANTIC Rule Construction…
                   OUTLOOK
HUMIDITY


                              Cloudy
 Not_H             Rain
           Humid
                          Sunny




                   CHECK: minCasesPerRule

  Cool
           Hot

   Mild


TEMPERATURE        Not_W
                                  Wind


                           WIND
FRANTIC Rule Construction…
                        OUTLOOK
HUMIDITY


 Not_H                  X
                       Rain       X
                                 Cloudy

           Humid




  Cool
           Hot

   Mild
                   CHECK!

TEMPERATURE             Not_W
                                  Wind


                                WIND
FRANTIC Rule Construction…
                      OUTLOOK
HUMIDITY


 Not_H                X
                      Rain      X
                               Cloudy

           Humid




             CHECK!

  Cool
           Hot

   Mild


TEMPERATURE
                        X
                      Not_W
                                Wind


                              WIND
    IRL – Training Set Adjustment
   Removal of training examples
   Re-weighting of training examples based on
    current best rule (class-independent IRL,
    Hoffmann 2004)
   Use of indicators for cooperation/competition
    between current rule and rules already in rule
    base (class-dependent IRL, Gonzales & Perez
    1999)
Classification Accuracy…
Number of Rules…
 minCasesPerRule Robustness…

Saturday Morning dataset – predictive accuracy while
                varying parameter
minCasesPerRule Robustness…

   Iris dataset – predictive accuracy while
               varying parameter
                  Future Work

   Identify and analyse parameter interactions

   Investigate impact of training adjustment method
    on parameter robustness

   Devise, explore and compare alternative
    approaches to training set adjustment

   Deepen understanding of IRL strategy by
    comparing different rule discovery mechanisms
      Encouraging Complementary
          Fuzzy Rules within
        Iterative Rule Learning

  Michelle Galea                     Qiang Shen
 School of Informatics      Department of Computer Science
University of Edinburgh           University of Wales
    Edinburgh, UK                  Aberystwyth, UK

                      Vishal Singh
               Larson & Toubro, EmSys Ltd.
                     Bangalore, India