Transcript Inductive Learning in Design: A Method and Case Study Concerning

Inductive Learning in Design:
A Method and Case Study Concerning Design of
Antifriction Bearing Systems
Machine Learning and Data Mining : Methods and
Applications
1999년 6월 19일 토요일
99406-810 산업공학과
허원창
Contents
 Introduction
 Exemplary problem
 Testing and Training events
 Exemplary rule set obtained
 Empirical errors of learned rule set
 Degree of Confidence
 Conclusion
Introduction
 A Method for Learning Design Rule
– in design process - design knowledge is important but
ambiguous, and there are many solutions in design problem
– in applying Inductive Learning Method - recognizing design
knowledge and representing it in the for of rule is important
– in this chapter - learning rules for selecting anti-friction
bearing systems
 Global Steps
–
–
–
–
defines attributes used for characterizing design examples
describe design examples with selected attributes
determining training and testing examples
learning through AQ15c and obtaining rule set
Example Problem
 Design of Bearing arrangement
 Design Process
Training and Testing Events
 Design Knowledge Source
– catalogues of rolling bearing, text books on machine design,
special publications issued by producers of bearing.....
– Conversions of quantitative data to qualitative data
 Database Examples
– bearing types : deep grove ball bearing, angular contact ball
bearing, self-aligning ball bearing, cylindrical roller bearings..
– 10-26 events for each bearings
– 101088 possible events
– need more events from design experts
Domains of Attributes
 Domains of Attributes
Exemplary Training Events
 training events of the class ‘deep groove ball bearing’
Exemplary rules
# of unique events that support rule
total # of events that support rule
 exemplary rule concerning ‘deep groove ball bearing’
Empirical Error of learned rule sets
 overall empirical error rate
Eov 
number of errors
number of testing events
 Empirical omission error rate
E om
1 n k
number of omission errors for class k
k

E om , E om

n k 1
number of positive examples for class k

 Empirical comission error rate
E cm
1 n k
number of comission errors for class k
k

E cm , E cm

n k 1
number of negative examples for class k

Testing Results
 Testing results using ‘leave-one-out’ method
Evaluation of Training Example
 Evaluation of training example
Exemplary Degree of Confidence
 exemplary Degree of confidence
Conclusion
 In problems of deriving useful design knowledge in
order to aid designer in routine design task
– The feasibility of the application of machine learning in case
of selecting the type of bearing.
– can suggests several solution to designers.
– The ruleset obtained features high degree of accuracy.
– Further verification of results require cooperation with skilled
designers