A Method for Temporal Knowledge Conversion G. Guimarães 1 and A. Ultsch 2 1 CENTRIA, U. Nova de Lisboa and Department of Mathematics, University of Évora Portugal 2 Department of Mathematics and Computer Science Philipps University of Marburg Germany Abstract In this paper we present a new method for temporal knowledge conversion, called TCon. The main aim of our approach is to perform a transition, i.e. conversion, of temporal complex patterns in multivariate time series to a linguistic, for human beings understandable description of the patterns. The main idea for the detection of those complex patterns lies in breaking down a highly structured and complex problem into several subtasks. Therefore, several abstraction levels have been introduced where at each level temporal complex patterns are detected successively using exploratory methods, namely unsupervised neural networks together with special visualization techniques. At each level, temporal grammatical rules are extracted. The method TCon was applied to a problem from medicine, sleep apnea. It is a hard problem since quite different patterns may occur, even for the same patient, as well as the duration of each pattern may differ strongly. Altogether, all patterns have been detected and a meaningfull description of the patterns was generated. Even some kind of “new” knowledge was found. 1. Introduction In recent years there has been an increasing development towards more powerfull computers, such that nowadays a great amount of data from, for example, industrial processes or medical applications, is gathered. These measured data are often said to be a starting point for an enhanced diagnosis or control of the underlying process. Particularly interesting for handling noisy or inconsistent data are artificial neural networks (ANN). On the other side, systems with traditional artificial intelligence (AI) technologies have been successful in areas like diagnosis, control and planing. The advantages of both technologies are wide-ranging. However, the limits of these approaches, namely the incapacity of ANN to explain their behaviour and on the other hand, the acquisition of knowledge for AI systems, are important problems to be adressed. Recently, there has been an increased interest in hybrid systems that integrate AI technologies and ANN to solve this kind of problems [2]. It its worth to remark here that essentially hybrid systems have been developed that entail several modules, each implemented in a different technology, and that cooperate with another. In contrast, we are mainly interested in hybrid systems that perform a knowledge conversion, i.e. a transition between distinct knowledge representation forms [14]. A symbolic knowledge representation of a subject should always be in a linguistic, for human beings understandable form. Examples for linguistic representation forms are natural languages, as German or English, but also predicate logic, mathematical calculus, etc. In contrast, a subsymbolic knowledge representation always entails numerous