Novel Approaches to Optimised Self-configuration in High Performance Multiple Experts M.C....
-
Upload
mervin-atkinson -
Category
Documents
-
view
220 -
download
0
description
Transcript of Novel Approaches to Optimised Self-configuration in High Performance Multiple Experts M.C....
Novel Approaches to Optimised Self-configuration in High
Performance Multiple Experts
M.C. Fairhurst and S. HoqueUniversity of Kent
UK
A. F. R. Rahman BCL Technologies Inc.
USA
Basic Problem Statement
• Given a number of experts working on the same problem, is group decision superior to individual decisions?
Is Democracy the answer?
• Infinite Number of Experts• Each Expert Should be Competent
How Does It Relate to Character Recognition?
Each Expert has its:• Strengths and Weaknesses• Peculiarities• Fresh Approach to Feature Extraction• Fresh Approach to Classification• But NOT 100% Correct!
Practical Resource Constraints
Unfortunately, We Have Limited• Number of Experts• Number of Training Samples• Feature Size• Classification Time• Memory Size
Solution
• Clever Algorithms to Exploit Experts– Complimentary Information– Redundancy: Check and Balance– Simultaneous Use of Arbitrary Features and
Classification Routines
How are they Employed?
Expert1 Expert 2 Expert n
Horizontal Systems
How are they Employed?
Vertical Systems
Expert 1
Expert 2
Expert n
How are they Employed?
• Combined System:– A hybrid of Horizontal and
VBertical– More Complicated to
Analyse?– Even more Complicated to
Optimise?
What to Optimise?
• Number of Experts in a configuration• Type of Expert in each Position in the
hierarchy• Optimising Criteria
– Do we want a fast system? Or– Do we want an accurate System?
Proposed Methodology
• Genetic Algorithm: A Generalised Search and Optimisation Method
• Problem Coding:– Chromosome Structure– Fitness Function– Genetic Operators
Methodology • Chromosome Structure: A
Classifier is a Machine Obeying a Set of Production Rules. A Generalised Rule is:<classifier>::=<condition>:<message>– <condition> part is a pattern matching
device– <message> part is a feedback
mechanism
Methodology
• Fitness Function: Fitness = Correct_Patterns/Total_Patterns
• Correct_Patterns corresponds to the number of correctly identified patterns in one cycle
• Total_Patterns corresponds to the number of total patterns being fed to the optimising process
Methodology • Genetic Operators:
– Reproduction: • Weighted Roulette Wheel (Goldberg)• Stochastic Remainder Selection (Booker)• Tournament Selection (Brindle)
– Crossover: Swapping at [1,l-1]– Mutation: Random variation
• Single gene only
Selection of a Specific Problem
Expert 1
Expert 2 Expert 3
Expert 4
Decision Compilation
Selection of a Database• Machine Printed Characters Extracted from
British Envelopes• Collected Off-line• Total 34 Classes (0-9, A-Z, no Distinction
between 0/O and I/1)• Total Samples of Over 10,200 characters• Size Normalised to 16X24
Performance of the Classifiers
Classifiers % Error
BWS 1.76
FWS 1.52
MPC 3.90
MLP 1.66
Performance of the Combination
Classifier Position % Error
BWS 1
FWS 4
MPC 3
MLP 2
1.03
The Optimised Combination
Classifier Position % Error
BWS Unused
FWS 2
MPC 1
MLP 3
0.92
Generality of the Solution: Generation of a Vertical System
Expert 1
Expert 2
Expert 4
Decision Compilation
Input Pattern
Classification Decision
Expert 1
Expert 2
Expert 3
Input Pattern
Classification Decision
Optimization for the Vertical System
Optimized Parameters
BWS Sub-set size
FWS Sub-set size
MPC Sub-set size
MLP Sub-set size
2 10 4 8 1 5 3 2
Combined % Error: 1.01
Generality of the Solution: Generation of a Horizontal System
Expert 2 Expert 3
Expert 4
Decision Compilation
Input Pattern
Classification Decision
Expert 1 Expert 2
Expert 3
Decision Compilation
Input pattern
Decision Combination
Classification Decision
Optimization for the Horizontal System
Optimized Parameter
BWS FWS MPC MLP Error %
Weighting Factor
0.14 0.53 0.11 0.22 0.92
Conclusion• Multiple Expert Solutions can be made more
Robust by optimising these structures• Optimisation is made with GA approach• The adopted multiple expert configuration is
generic: it can produce both vertical and horizontal systems (in addition to the hybrid system)
• The optimization approach is generic: it man optimize both vertical and horizontal systems (in addition to the hybrid system)