01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 Elsevier AMS Job code: BFN Ch05-N52710 23-1-2007 8:05 a.m. Page:149 Trimsize:165×240 MM Quantitative Structure-Activity Relationships (QSAR) for Pesticide Regulatory Purposes 149 Edited by Emilio Benfenati © 2007 Elsevier B.V. All rights reserved. Chapter 5 Hybrid systems Nicolas Amaury 1 , Emilio Benfenati 2 , Severin Bumbaru 3 , Antonio Chana 2 , Marian Craciun 3 , Jacques R. Chrétien 1 , Giuseppina Gini 4 , Gongde Guo 5 , Frank Lemke 6 , Viorel Minzu 3 , Johann-Adolf Müller 6 , Daniel Neagu 5 , Marco Pintore 1 , Silviu Augustin Stroia 3 , Paul Trundle 5 1 BCX, Biochemics Consulting SAS, Orléans, France 2 Laboratory of Environmental Chemistry and Toxicology, Istituto di Ricerche Farmacologiche “Mario Negri”. Milano, Italy 3 Department of Computer Science and Engineering, University “Dunarea de Jos”, Galati, Romania 4 Department of Electronic and Information, Politecnico di Milano, Milano, Italy 5 Department of Computing, School of Informatics, University of Bradford, Bradford, UK 6 KnowledgeMiner, Berlin, Germany 1. INTRODUCTION: GOALS OF THE HYBRID SYSTEMS The term ‘hybrid system’ dates back to the development of expert systems (Jackson, 1999) from the initial rule-based aspect to the modern modeling and interpretation systems. Most of the accent in the beginning has been on the idea of making use of more representations of the problem, more paradigms of knowledge representation, and more algorithms to find a solution. A seminal work by Gallant (1993) introduced a way to look together to neural networks and rule-based systems. In his approach, a net, built from data and in absence of symbolic knowledge, is used to extract rules. This idea developed in the artificial intelligence (AI) community the well-known area of integrating connectionist and symbolic systems. The starting machine learning community developed in the same years another way to make use of data in the absence of knowledge that led to the development of inductive trees, well exemplified by C4.5 (Quinlan, 1993). Integrating different representations and solutions is a direction taken in AI in the years around 1995. The term ‘expert system’ in those years was almost replaced by the term ‘intelligent system’ or ‘intelligent agent’. Using different representations to reach a common agreement or a problem solution led to the idea of using computational different methods on different problem representations, to make use of their relative strengths. Examples are the