Selection of optimal physical activity classification model using measurements of accelerometer

The paper solves the problem of selecting optimal stable models for classification of physical activity. Each type of physical activity of a particular person is described by a set of features generated from an accelerometer time series. In conditions of feature's multicollinearity, selection of stable models is hampered by the need to evaluate a large number of parameters of these models. Evaluation of optimal parameter values is also difficult due to the fact that the error function has a large number of local minima in the parameter space. In the paper, the optimal models from the class of two-layer artificial neural networks are chosen. The problem of finding the Pareto optimal front of the set of models is solved. The paper presents a stepwise strategy of building optimal stable models. The strategy includes steps of deleting and adding parameters, criteria of pruning and growing the model and criteria of breaking the process of building. The computational experiment compares the models generated by the proposed strategy on three quality criteria — complexity, accuracy, and stability.