Fuzzy Rule Reduction using Sparse Auto-Encoders

Fuzzy Rule based regression, classification and control have found great use in modern applications due to its simplicity, flexibility and capability. A key issue in all such methods is the computation time. Computational complexity of training and testing is linearly dependent on the size of fuzzy rule base and the respective fuzzy rule space is exponentially dependent on data dimensionality. Sparse Auto-Encoders (SAs) have become popular in giving compact feature representations for image, audio and speech data and have helped in giving state of the art pattern recognition performances in most of the domains. These feature representation are learnt in an unsupervised fashion and are found to give higher order building blocks with which the data is seemingly made of. This paper proposes a method where SAs are used for getting compact feature representation of input data and if needed with reduced dimensionality. The regular fuzzy rule based models are then learnt from data in the new feature space. The method was tested for Regression and Classification problems, giving impressive results in both. The method with Regression problem gave comparable performance with almost half the number of rules and with Classification problem it gave improvement in classification accuracy by 2.67% while reducing the size of fuzzy rule base by 11.25 times and 7.5 times by number.