Direct calculation of predictions for K29/K29*

Implementation of the K29 learning algorithm presents the problem of finding a root of an often non-invertible function. Numerical methods giving approximations to these roots have small numerical inaccuracies. These inaccuracies, despite possibly seeming negligible, can accumulate quickly when applied to K29. The main mathematical result of this paper presents a simple but novel derivation of two formulae, which directly calculate predictions for the K29 (and the regularised version K29*) learning algorithms. We present comparisons between this new implementation and the numerical method through empirical investigation.