Trusted learner: An improved algorithm for trusted incremental function approximation
Andreas Buschermoehle, Jan Schoenke, Werner Brockmann · 2011
The complexity of technical systems increases drastically if they are operated in non-stationary or uncertain environments. Hence self-tuning, self-optimization and learnability get more and more important. At this as a way of achieving an intelligent system behavior, online learning systems are needed to overcome a possibly poor parameterization at engineering time and to adapt to new situations. But an online learning system has possibly uncertain knowledge at several stages of its learning process because of conflicting or sparse data. Thus it is crucial to reflect these uncertainties explicitly. In this paper we propose a method to accompany the learned knowledge by a so-called trust signal reflecting its trustworthiness. This meta-information can be exploited in the further system context. Additionally, as the specific focus here, it can also be used to steer and accelerate the learning process. Several examples show the benefits for learning from scratch and the new expressiveness in case of uncertain or changing target functions both on simulated and on real data.