Nusser S. Robust learning in safety-related domains: machine learning methods for solving safety-related application problems

Otto von Guericke University Magdeburg, 2009. — 140 p.Today, machine learning methods are successfully deployed in a wide range of applications. A multitude of different learning algorithms has been developed in order to solve classification and regression problems. These common machine learning approaches are regarded with suspicion by domain experts in safety-related application fields because it is often infeasible to sufficiently interpret and validate the learned solutions. Especially for safety-related applications, it is imperative to guarantee that the learned solution is correct and fulfills all given requirements. The basic idea of the approaches proposed within this thesis is to solve high-dimensional application problems by an ensemble of simple submodels, each of which is allowed to only use two or three dimensions of the complete input space. The restriction of the dimensionality of the submodels allows the visualization of the learned models. Thus a visual interpretation and validation according to the existing domain knowledge becomes feasible. Due to the visualization, an unintended and possibly undesired extra- and interpolation behavior can be discovered and avoided by changing the model parameters or selecting other submodels. Since the learned submodels are interpretable the correctness of the learned solution can therefore be guaranteed. The ensemble of the submodels compensates for the limited dimensionality of the individual submodels. The proposed ensemble methods are successfully applied on common benchmark problems as well as on real-world application problems with very high requirements on the functional safety of the learned solution.