Pergamon Computers them. Engng Vol. 22, Suppl., pp. S991-SIOOO, 1998 8 1998 Elsevier Science Ltd. All tights reserved Printed in Great Britain PII: SOO9?3-1354(98)00199-9 0098-1354/98 $19.00 + 0.00 Hybrid, fuzzy and neural adaptation in case- based reasoning system for process equipment selection T. Koiranen, T. Virkki-Hatakka, A. Kraslawski*, L. Nystriim Lappeenranta University of Technology, Dept of Chemical Technology, POB 20, FIN 53851 Lappeenranta, Finland *FAX: +358 5 621 2199 ABSTRACT E-MALL: Andrzej.Krasiawski@lut.fi Adaptation is the most difficult step in case-based reasoning. There are no general methods for the modification of the retrieved cases to fit the actual design problem. The use of new, context independent technique is an important research subject. In the paper, a hybrid adaptation system is proposed. Its main components are based on fuzzy logic and neural networks. The applicability of the proposed adaptation methods is examined in preliminary design of mixing system 0 1998 Elsevier Science Ltd. All rights reserved. INTRODUCTION Case-based reasoning (CBR) is based on the retrieval and adaptation of the old solutions to the new problems. There are five steps in CBR: introduction of a new problem, retrieval of the most similar cases, adaptation of the most similar solutions, validation of the current solution and system learning by adding the verified solution to the database of cases. There are several applications of CBR in engineering design, Mille et al (1995), Surma and Braunschweig (19%) and Vi-Hatakka et al. (1997). The benefits as well as the problems in process engineering design resulting Erom the use of CBR have been presented in Kraslawski et al (1995). However, the use of CBR in engineering disciplines has not yet reached the status suggested by the common use of experience in the solving of design problems. One of the reasons is the problem encountered in the adaptation phase of CBR. There are several techniques applicable in adaptation: constraint satisfaction, rule-based, simulation, causal diagnose and repair, etc., Maher et al. (1997). In the presented system, the use of neural nehvorks for adaptation in CBR is tested in preliminary design of equipment. SYSTEM DESCRIPTION The system real&s the main phases of CBR and its architecture is based on the programme presented by Virkki-Hatakka et al. (1997). A new method of adaptation based on neural networks is added in the actual version of the programme. The system is composed of two main parts. The akvefopmentpart consists of the mechanism for building of the cases’ description, presentation and storage. In the application part, there are implemented the main elements of case-based reasoning: introduction of a new s991 problem, retrieval of the most similar cases, adaptation, and storage. Adapt&n bawd on fuzqv logic (XL) The adaption is performed using the membership functions of the specific parameters. The membership functions have been built based on the values of the respective parameters in the retrieved cases. The details concerning the construction principles and the forms of the function have been presented in Viii-Hatakka et al. (1997). Adaptation based on an art@cial naml network (ANN) Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF) networks were applied in adaptation. MLP with one hidden layer was tmined using won with Lever&erg-Marquardt optimization method. ~emaindifficultyinusingMLPhasbeentheneedofa large amount of data to train the net. The lack of the sufftcient amount of design cases resulted in very poor quality of the obtained solutions. Next, RBF has been tested, and finally it has been selected as the most suitable for the adaptation phase in equipment design. The RBF network has a similar form to the MLP as it is a multi-layer, feed-fonuard network. However, the hidden units in the RBF are different from the units in the input and output layers. They contain the Radial Basis Function that is based on a Gaussian distribution. In the hidden layer of RBF, each hidden unit takes, as its input, all the outputs of the input layer x+ The hidden unit contains Radial Basis Function which has the parameters “centre” c and “width” a _ The centre of the basis function is a vector of numbers ci of the same size as the inputs to the unit and there is normally a diiatent centte c,,for each unit in the neural network. The radial basis function for the hth node of the hidden layer has the following form: