Encouraging Complementary Fuzzy Rules within Iterative Rule Learning

1
Encouraging Complementary Fuzzy Rules within
Iterative Rule Learning
2
Motivation

• Gain deeper understanding of IRL strategy for
  fuzzy rule base induction
• Test ACO as rule discovery mechanism within IRL

3
IRL Iterative Rule Learning
SPBA1
SPBA2
. . .
4
Ant Colony Optimisation The Basics
Constructionist, iterative algorithm

• Problem representation
• Probabilistic transition rule
• Local heuristic
• Constraint satisfaction method
• Fitness function
• Pheromone updating strategy

5
ACO for Fuzzy Rule Induction
ACO 1
. . . . . . . . . .
6
FRANTIC Rule Construction
OUTLOOK
HUMIDITY
Cloudy
Rain
Not
_
H
Humid
Sunny
Cool
Hot
Mild
TEMPERATURE
Not
_
W
Wind
WIND
7
FRANTIC Rule Construction
OUTLOOK
HUMIDITY
Cloudy
Rain
Not
_
H
Humid
Sunny
CHECK minCasesPerRule
Cool
Hot
Mild
TEMPERATURE
Not
_
W
Wind
WIND
8
FRANTIC Rule Construction
OUTLOOK
HUMIDITY
X
Cloudy
X
Rain
Not
_
H
Humid
Cool
Hot
Mild
CHECK!
TEMPERATURE
Not
_
W
Wind
WIND
9
FRANTIC Rule Construction
OUTLOOK
HUMIDITY
X
Cloudy
X
Rain
Not
_
H
Humid
CHECK!
Cool
Hot
S
u
n
n
y
Mild
X
TEMPERATURE
Not
_
W
Wind
WIND
10
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)

11
Classification Accuracy
12
Number of Rules
13
minCasesPerRule Robustness
Saturday Morning dataset predictive accuracy
while varying parameter
14
minCasesPerRule Robustness
Iris dataset predictive accuracy while varying
parameter
15
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

16
Encouraging Complementary Fuzzy Rules within
Iterative Rule Learning