Abstract

Virtual Stochastic Sensors (VSSs) aim to provide insight into stochastic processes by producing statistically relevant estimates of non-measurable system properties. During behavior reconstruction of these discrete stochastic systems the internal system state changes are often described as time-homogeneous distribution functions, as in Conversive Hidden non-Markovian Models (CHnMMs). However, the system behavior might change over time or the sample, used for the model creation, might not describe the system accurately. We have shown that detecting these changes is possible, yet the resource consumption for the re-estimation of the model was a clear problem. In this paper we present a solution to that problem by replacing the used statistical tests with Kernel Density Estimation (KDE) and by integrating the hidden model description into the proxel-based state space simulation method. By using the Change Adaptation Algorithm (CAA) this paper shows that adapting to runtime changes is possible, while preserving parameters on transitions where no change occurs. The algorithm was tested with 5 different types of Probability Density Functions (PDFs) which showed accurate results. By using the CAA one is able to construct adaptive models for behavior reconstruction without the need to fully parametrize the model. In this way, loss of modeling accuracy in the model construction process can be significantly decreased.