On-line Learning Based on Adaptive Similarity and Fixed Size Rule Base Joana Matos Dias 1 , António Dourado Correia 2 1 Research Student, 2 Professor CISUC-Centro de Informática e Sistemas da Universidade de Coimbra Departamento de Engenharia Informática PÓLO II da Universidade de Coimbra -Pinhal de Marrocos P 3030 COIMBRA PORTUGAL Phone: +351 39 7 000 000 Fax: +351 39 701 266 e-mail: joana@student.dei.uc.pt , dourado@dei.uc.pt Abstract: In this paper a methodology is developed to control linear and non- linear processes using a fuzzy approach with the main assumption that the output of the process is monotone with respect to the input. Beginning with an empty rule base, a fuzzy model is on-line built. The rule base has a fixed number of rules determined à priori and not depending on the complexity of the process. The controller experiences a learning phase during which it learns how to control the process, that is repeated whenever there is some change in the process behaviour. The inference and defuzzification mechanisms have their background on the Fuzzy Equality Relations Theory, using an adaptive degree of similarity. The proposed controller was successfully applied in simulation for linear and non-linear systems and practical essays were made on a real non- linear thermal process, for both the regulation and the tracking problem. 1 Introduction Modern industrial processes are becoming each time more and more complex. Most of them have non-linear features partially unknown that make the use of mathematical models difficult and sometimes inefficient. Neural network controllers try to overcome this problem, functioning as a black box that is trained to control the process apparently without knowing much about its dynamics. But if the process modifies its behaviour in time the results will not be satisfactory and the network will have to be trained again. In fuzzy controllers, one has the opportunity to include the operator’s knowledge in the controller algorithm in a non-mathematical way. But the time spent trying to tune the parameters of the controller correctly can be unbearable. The proposed controller is a fuzzy controller that does not need the expert knowledge about the process. Its rule base is empty at first, and is constructed based on the values the controller gets from and sends to the process. The construction of the rule-base is on-line and is a never ending process, so that the controller can react whenever there is a change in the process behaviour. When the rule base is empty, or if there is a change