Representing, adapting and reasoning with uncertain, imprecise and vague information G. Vouros * Department of Mathematics, University of the Aegean, Karlovassi Samos 83200, Greece Abstract This paper presents a formulation for representing and reasoning with uncertain knowledge towards making diagnoses in cases where there is no deep knowledge on the mechanisms that make a set of observations occur, in cases where knowledge about these mechanisms is based on simplifying assumptions or the representation of the underlying mechanisms and their exploitation by a system are resource demanding. Towards this aim, the paper describes a formulation for representing and reasoning with uncertain knowledge. The aim is to provide a generic framework for the development of expert systems that provide assistance to humans during diagnosis. The paper describes in detail the nature of the diagnosis task as part of the exploration task, and identifies the types of uncertainty accommodated in this task. It presents the proposed formulation and provides simple examples to make the mathematical basis more comprehensive. The paper proposes a method for adapting uncertainties to known cases, towards fine-tuning the representation.  2000 Elsevier Science Ltd. All rights reserved. Keywords: Diagnosis; Uncertain knowledge; Causal diagrams; Learning 1. Introduction Diagnosis in cases where there is no “deep” knowledge on the mechanisms that make a set of “observations” occur, requires representing uncertain knowledge and reasoning with it. The objective of the task is to exploit these observa- tions and find the most accurate model of the situation on hand. For instance, to assess the characteristics, preferences and needs of a website visitor, a system needs to associate behavioral cues with features of user-models (Jameson, 1996). In another context, a web of medical doctors needs to acquire information about a patient, form “observations”, interpret them, diagnose the situation and predict the occur- rence of future events. Generally, such observations can be either features obtained by interpreting raw data from various origins (e.g. laboratory analyses, statistical data) or direct indications (e.g. hits on web pages, symptoms) observed with naked eye. The aim of our research is to develop a framework for the development of knowledge-based systems that exploit uncertain, imprecise and vague information and provide advice to humans for diagnosing a situation. As already indicated, diagnosis aims to find the most accurate model of a situation on hand. Models represent classes of entities that are of particular interest and comprise systematically arranged information describing the essential attributes of these classes. They include types of computer user, geologi- cal deposit, operational problems that may appear in a plant, etc. Borrowing terminology from other disciplines, a model can be descriptive, generic, or be formed by a combination of these model types. A descriptive model provides the essential features that a class of entities must possess. A genetic model represents “deep knowledge” about the interesting entities, relates the features of entity types among themselves and describes the exact mechanisms that make these features occur. The aim of genetic models is to bring coherence to the particular set of models’ features. This paper deals with the assessment of features of descriptive models. It presents a formulation for represent- ing and reasoning with uncertain knowledge towards making diagnoses in cases where there is no deep know- ledge on the mechanisms that make model features occur, in cases where knowledge about these mechanisms is based on simplifying assumptions or the representations of the under- lying mechanisms and their exploitation by a system are resource demanding. We assume that in these cases, human experts possess domain knowledge expressed in terms of observations and factors, related among themselves via causal dependencies. The role of domain knowledge during the diagnosis task is to provide alternative reasonable aggregations of observations Expert Systems with Applications 19 (2000) 167–192 PERGAMON Expert Systems with Applications 0957-4174/00/$ - see front matter  2000 Elsevier Science Ltd. All rights reserved. PII: S0957-4174(00)00031-2 www.elsevier.com/locate/eswa * Tel.: +30-273-82226; fax: +30-273-33896. E-mail address: georgev@aegean.gr (G. Vouros).