Optimization of Probabilistic Switching Models Based on a Two-Step Clustering Approach

This paper proposes a method for optimizing the clusters employed as discrete random variables in probabilistic switching models. The proposed optimization facilitates obtaining a low number of discrete components while guaranteeing a high performance at predicting and detecting abnormalities in time series of data. Our method is composed of a two-step clustering approach that first considers a partitional clustering to obtain an initial semantic representation of data. Then it performs a hierarchical clustering to decrease the number of clusters while preserving coherent models employed for predicting future time instances at the testing phase. Odometry data from a real vehicle that performs different tasks in a closed, controlled environment is used to evaluate the proposed method.