Abstract

Virtual Stochastic Sensors calculate statistically relevant estimates in indirectly observable discrete stochastic systems. This is done by the proxel-based analysis that aims to reconstruct the relevant part of the state space with an iterative process. Strategically removing non-relevant proxels from the analysis (pruning) to reduce runtime overhead might potentially affect the results. And while the impact on the decoding problem has already been analysed in detail, the effect on the evaluation problem was not yet discussed.

The paper discusses three pruning strategies and compares their properties in case of the evaluation task. The theoretical statements are empirically proven using a car rental agency model in form of a Conversive Hidden non-Markovian Model.

The results show that in case of well chosen parameters all three pruning strategies are able to reach the same evaluation probability. The major difference between the strategies is due to their runtime properties which need to be carefully aligned with the use-case to reach optimal behavior. Based on the results the fixed number of proxels pruning strategy provides highly predictable execution time, while the fixed threshold pruning is very good at discovering a broader spectrum of the state space. The variable pruning is a very good trade-off between the previous strategies enabling lower thresholds and thorough state space analysis while maintaining acceptable execution times at the cost of more complex parametrisation.