Noise elimination in approximation and time series prediction with hinging hyperplanes

This paper presents a method to approximate multidimensional functions using the Sweeping Hinge Algorithm (SHA) in combination with the Truncated Hinging Hyperplanes (THH) as approximation units. The paper focuses on the real learning problems where some noise is present in the available examples. We show how the method provides good approximations, due to the simplicity of the selected units and to the fact that the convergence properties of the SHA algorithm are not influenced by the noise present in the data. Fair simulations contribute to support the good behavior of this purely constructive method against the noise.