Industrial and Engineering Applications or Artificial Intelligence and Expert Systems

Takushi Tanaka, Setsuo Ohsuga, Ali Moonis, Takushi Tanaka, Setsuo Ohsuga, Ali Moonis

• Publisher: CRC Press
• Year: 2022
• eBook ISBN: 9781000674071
• Topic: Computer Science
• Subtopic: Computer Science General
• Index: Computer Science

This volume includes the proceedings from Proceedings of the Ninth International Conference Fukuoka, Japan, June 4-7, 1996. This work represents a broad spectrum of new ideas in the field of applied artificial intelligence and expert systems, and serves to disseminate information regarding intelligent methodologies and their implementation in solving various problems in industry and engineering.

Machine Learning

COIN CLASSIFICATION USING A NOVEL TECHNIQUE FOR LEARNINGCHARACTERISTIC DECISION TREES BY CONTROLLING THE DEGREE OF GENERALIZATION

Paul Davidsson

Department of Computer Science, Lund University

Lund, Sweden, S-221 00

Email: [email protected]

ABSTRACT

A novel method for learning characteristic decision trees is applied to the problem of learning the decision mechanism of coin-sorting machines. Decision trees constructed by ID3-like algorithms are unable to detect instances of categories not present in the set of training examples. instead of being rejected, such instances are assigned to one of the classes actually present in the training set. To solve this problem the algorithm must learn characteristic, rather than discriminative, category descriptions. In addition, the ability to control the degree of generalization is identified as an essential property of such algorithms. A novel method using the information about the statistical distribution of the feature values that can be extracted from the training examples is developed to meet these requirements. The central idea is to augment each leaf of the decision tree with a subtree that imposes further restrictions on the values of each feature in that leaf.

1. INTRODUCTION

One often ignored problem for a learning system is how to know when it encounters an instance of an unknown category. In many practical applications it cannot be assumed that every category is represented in the set of training examples (i.e., they are open domains [Hut94]) and sometimes the cost of a misclassification is too high. What is needed in such situations is the ability to reject instances of categories that the system has not been trained on.

In this article we will concentrate on an application concerning a coin-sorting machine of the kind often used in bank offices.¹ Its task is to accept and sort (and count)a limited number of different types of coins (for instance, a particular country’s), and to reject all other coins. The vital part of the machine is a sophisticated sensor that the coins pass one by one. The sensor measures electronically five properties (diameter, thickness, permeability, and two kinds of conductivity)of each coin, which all are given a numerical value. Based on these measurements the machine decides of which type the current coin is: if it is of a known type of coin, it is sorted, otherwise it is regarded as an unknown type and is rejected.

The present procedure for constructing the decision mechanism is carried out mostly by hand. A number of coins of each type are passed through a sensor and the measurements are recorded. The measurements are then analyzed manually by an engineer, who chose a minimum and a maximum limit for each property of each type of coin. Finally, these limits are loaded into the memory of the machine. When the machine is about to sort a new coin, it uses the limits in the following way: if the measurement is higher than the minimum limit and lower than the maximum limit for all properties of some type of coin, the coin is classified as a coin of this type. If this is not true for any known type of coin, the coin is rejected.

Thus, in the present method, which is both complicated and time consuming, it is the skill of the engineer that decides the classification performance of the coin-sorting machine. Moreover, this procedure must be carried out for every new set of machines (e.g., for each country’s). In addition, there are updating problems when a new kind of coin is introduced. This is not only applicable when a new denomination is introduced, or when the appearance of an old denomination is changed. It is, for instance, not unusual that the composition of the alloy is changed. In fact, this happens often undeliberately as it is difficult to get exactly the same composition every time and, moreover, there are sometimes trace elements of other metals in the cauldron. Another kind of problems comes from the fact that it is difficult to make all the sensors exactly alike. As a consequence of this and the fact that all machines used for the same set of coins use the same limits, each sensor must be calibrated. Moreover, this calibration is not always sufficient and service agents must sometimes be sent out to adjust the limits on particular machines.

Figure 1 Discriminative (left) Versus Characteristic (right) Category Descriptions

However, if we construct a method which automatically learns the decision mechanism, it would be possible to bring down the effects of these problems. In short, the task to be so...

Table of contents

• Cover
• Half Title
• Title Page
• Copyright Page
• Table of Contents
• Preface
• Conference Organization
• List of Sponsors
• Invited Papers
• Automated Reasoning
• CAD/CAM
• Case-Based Reasoning
• Database
• Decision Support
• Diagnosis
• Distributed Ai
• Fuzzy Logic And Control
• Genetic Algorithm
• Intelligent Tutoring
• Knowledge Acquisition
• Knowledge-Based Systems
• Knowledge Representation
• Logic Programming
• Machine Learning
• Manufacturing
• Monitoring
• Neural Network
• Neural Network Applications
• Natural Language
• Planning and Scheduling
• Practical Applications
• Robotics
• Vision
• Abstracts for Poster Session
• Index of Authors