| Abstract |
Artificial intelligence (AI) is now experiencing extraordinary growth, and applications of its ideas and methods are appearing in many fields. Among its most visible and important successes are the development of expert systems. In this context, it is important to ask what the limitations of the current methods are and what new directions research in this field should take. One of the obvious limitations relates to machine learning. Current AI systems have very limited learning abilities or none at all. All of their knowledge must be programmed into them. When they contain an error, they cannot correct it on their own. They will repeat it endlessly, no matter how many times the procedure is executed. Generally speaking, these systems lack the ability to draw inductive inferences from information given to them. The ability to classify objects or events as members of known classes is a very common task for learning systems. A well known approach to heuristic classification is decision tree induction. An alternative to decision tree induction is given by the well known algorithm id3. id3 uses a heuristic search process to find a set of discriminant descriptions between classes, given: (1) A set of observational statements each of which is assigned to a certain class and (2) a universe of classes. Working in a recursive manner, the algorithm selects the most discriminant attribute by maximizing an information gain function at each step. The result is a tree in which nodes represent tests on attributes, while branches are possible values of the corresponding attributes. id3 does not take into account any background information resulting to a set of rules that are far from the expert's model. In this work, we present iddd, which extends id3 by using dependency relationships, between attributes and/or attribute value sets, as domain knowledge agents. iddd is based on NewId, an enhanced implementation of id3, developed by the Turing Institute. The basic premise of iddd lies in the deployment of domain knowledge in the decision tree induction process. We introduce dependency relationships between attributes that are provide some structure over the rather `flat' data representation used by NewId itself. The attribute that depends on another attribute is called Daughter while the latter attribute is called Mother. A simple dependency relationship states that information represented by attribute Daughter should be useful only when it is combined with information represented by attribute Mother. Another form of the relation defined above is the exclusion of an attribute when a specific value has already been assigned to another attribute. We call this dependency exclusive. We should note here that any Daughter attribute may have only one mother, while this restriction does not apply to a mother attribute. More often than not, NewId is able to `handle' relationships such as the ones defined above, yet this is achieved in an implicit manner and it is not based on explicit modeling of domain knowledge. However, when the value of a mother attribute is unknown, NewId may drift into irrelevant attribute selection and splitting. Furthermore, when two attributes A and B score equally in information gain, NewId may select B instead of A. The outcome is a set of rules which may be accurate, from the point of view of classification accuracy but meaningless from the expert's point of view. The effectiveness of iddd is demonstrated through carefully designed experiments involving a medical application. The effectiveness of dependency relationships defined in iddd indicates that further work is necessary towards the investigation, formal definition, and handling of domain knowledge in the inductive process.
|
Χρήση Γνώσης του Πεδίου Εφαρμογής για την Ενίσχυση των Επαγωγικών Μηχανισμών Μάθησης Μέσω Παραδειγμάτων
