Learning Systems for Nonlinear Mapping

Abstract

The many complex problems facing researchers and engineers demand innovative solutions. Machine learning techniques are growing in popularity due to their versatility and power. However, challenges remain. Popular machine learning algorithms such as Artificial Neural Networks are difficult to train, and require many designer choices that heavily impact the performance of the network. Furthermore, the randomized starting point of most ANN variants means that even if optimal choices are made, it may still take multiple trials to obtain satisfactory results. Fuzzy Systems are also widely used, but cannot tackle high dimensional problems or produce outputs of similar quality to neural networks. A novel defuzzification routine based on cubic splines seeking to improve the performance of FS is introduced, and compared to many state of the art machine learning techniques. The experimental results show the proposed algorithm performs competitively with popular machine learning methods, while not requiring a lengthy training process.