A methodology using fuzzy logic to optimize feedforward artificial neural network configurations

After a problem has been formulated for solution by using artificial neural network technology, the next step is to determine the appropriate network configuration to be used in achieving a desired level of performance. Due to the real world environment and implementation constraints, different problems require different evaluation criteria such as: accuracy, training time, sensitivity, and the number of neurons used. Tradeoffs exist between these measures, and compromises are needed in order to achieve an acceptable network design. This paper presents a method using fuzzy logic techniques to adapt the current network configuration to one which is close to (if not at) the optimal configuration. The fuzzy logic provides a method of systematically changing the network configuration while simultaneously considering all of the evaluation criteria. The optimal configuration is determined by a cost function based on the evaluation criteria. The proposed methodology is applied to an elementary classifier network as an illustration. The procedure is then used to automatically configure a network used to detect incipient faults in an induction motor as a real world application