AN INTERPOLATION TECHNIQUE FOR LEARNING WITH SPARSE DATA

摘要

The occurence of sparse data is a stringent problem for Black Box system identification. The lack of measurement data leads to a high modeling error in the resulting gaps. There are a number of techniques described in the literature to fill in these gaps. Some of these techniques are only useable for a low input count (splines). Others suffer from noisy data (interpolation), severe parameter biasing (adding input noise) or time-consuming recursive calculations (Errors-In-Variables). This paper describes a linear feedforward interpolation technique that uses probability techniques, based on 95% or 99% certainty bounds. A smoothing factor can be used to trade off bias and variance, based upon probability theory.