I.P. Vlahavas and C.D. Spyropoulos (Eds.): SETN 2002, LNAI 2308, pp. 30–41, 2002. © Springer-Verlag Berlin Heidelberg 2002 Multi-inference with Multi-neurules Ioannis Hatzilygeroudis and Jim Prentzas University of Patras, School of Engineering Dept of Computer Engin. & Informatics, 26500 Patras, Hellas (Greece) & Computer Technology Institute, P.O. Box 1122, 26110 Patras, Hellas (Greece) {ihatz, prentzas}@ceid.upatras.gr Abstract. Neurules are a type of hybrid rules combining a symbolic and a connectionist representation. There are two disadvantages of neurules. The first is that the created neurule bases usually contain multiple representations of the same piece of knowledge. Also, the inference mechanism is rather connectionism oriented than symbolism oriented, thus reducing naturalness. To remedy these deficiencies, we introduce an extension to neurules, called multi- neurules, and an alternative inference process, which is rather symbolism oriented. Experimental results comparing the two inference processes are also presented. 1 Introduction There have been efforts at combining the expert systems approach and the neural networks (connectionism) one into hybrid systems, in order to exploit their benefits [1]. In some of them, called embedded systems, a neural network is used in the inference engine of an expert system. For example, in NEULA [2] a neural network selects the next rule to fire. Also, LAM [1] uses two neural networks as partial problem solvers. However, the inference process in those systems, although gains efficiency, lacks the naturalness of the symbolic component. This is so, because pre- eminence is given to the connectionist framework. On the other hand, connectionist expert systems are integrated systems that represent relationships between concepts, considered as nodes in a neural network. Weights are set in a way that makes the network infer correctly. The system in [3] is a popular such system, whose inference engine is called MACIE. Two characteristics of MACIE are: its ability to reason from partial data and its ability to provide explanations in the form of if-then rules. However, its inference process lacks naturalness. Again, this is due to the pure connectionist inference approach. In a previous work [4], we introduced neurules, a hybrid rule-based representation scheme integrating symbolic rules with neurocomputing, which gives pre-eminence to the symbolic component. Thus, neurules give a more natural and efficient way of representing knowledge and making inferences. However, there are two disadvantages of neurules, from the symbolic point of view. Neurules are constructed