Fuzzy Logic Programming for Tuning Neural Networks

Wide datasets are usually used for training and validating neural networks, which can be later tuned in order to correct their final behaviors according to a few number of test cases proposed by users. In this paper we show how the FLOPER system developed in our research group is able to perform this last task after coding a neural network with a fuzzy logic language where program rules extend the classical notion of clause by including on their bodies both fuzzy connectives (useful for modeling activation functions of neurons) and truth degrees (associated with weights and biases in neural networks). We present an online tool which helps to select such operators and values in an automatic way, accomplishing with our recent technique for tuning this kind of fuzzy programs. Moreover, our experimental results reveal that our tool generates the choices that better fit user's preferences in a very efficient way and producing relevant improvements on tuned neural networks.