The paper discusses the problem of rule weight tuning in neuro-fuzzy systems with parameterized consequences in which rule weights and the activation of the rules are not interchangeable. Some heuristic methods of rule weight computation in neuro-fuzzy systems with a hierarchical input domain partition and parameterized consequences are proposed. Several heuristics with experimental results showing the advantage of their usage are presented.
2
Dostęp do pełnego tekstu na zewnętrznej witrynie WWW
Continuous attributes are usually discretized into intervals in machine learning and data mining. Our knowledge representation is, however, not always based on such discretization. For example, we usually use linguistic terms for dividing our ages into some categories with fuzzy boundaries. In this paper, we examine the effect of fuzzy discretization on the classification performance of fuzzy rule-based systems through computer simulations on simple numerical examples and real-world pattern classification problems. For exulting such computer simulations, we introduce a control parameter that specifies the overlap grade between adjacent antecedent fuzzy sets in fuzzy discretization. Interval discretization can be viewed as a special case of fuzzy discretization with no overlap. Computer simulations are performed using fuzzy discretization with various specifications of the overlap grade. Simulation results show that fuzzy rules have high generalization ability even when the domain interval of each continuous attribute is homogeneously partitioned into linguistic terms. On the other hand, generalization ability of rule-based systems strongly depends on the choice of theshold values in the case of interval discretization.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.