Design of controller for mixed data type - A composite fuzzy-neural approach

Input feature vectors in feed-forward neural network are generally processed by linear separation through plane or hyperplane. But discrete features are difficult to feed into these nonlinear computational processing elements. Even if the processing is carried out using some encoding mechanisms of quantization,retaining information of the original categorical variables appears to be difficult at times. This paper proposes a methodology to design an online controller through fuzzification of the discrete input features first and,then through supervised learning based on adaptive neuro-fuzzy inference system (ANFIS). Continuous feature spaces are processed simultaneously through multi-layered perceptrons (MLP) and both the systems are connected through a linear filter for prediction of response variables using log-likelihood cost function. This proposed architecture has been tested on standard data set for its efficacy. The system is proven to be more efficient as compared to the existing methods in terms of error metrics. The estimation is independent of the number of discrete input features and in case if the data set has several categorical features,this method will still be more effective than the available systems. © School of Engineering,Taylor’s University.