History-Dependent Stochastic Petri Nets

Stochastic Petri Nets are a useful and well-known tool for performance analysis. However, an implicit assumption in the different types of Stochastic Petri Nets is the Markov property. It is assumed that a choice in the Petri net only depends on the current state and not on earlier choices. For many real-life processes, choices made in the past can influence choices made later in the process. For example, taking one more iteration in a loop might increase the probability to leave the loop, etc. In this paper, we introduce a novel framework where probability distributions depend not only on the marking of the net, but also on the history of the net. We also describe a number of typical abstraction functions for capturing relevant aspects of the net's history and show how we can discover the probabilistic mechanism from event logs, i.e. real-life observations are used to learn relevant correlations. Finally, we present how our nets can be modelled and simulated using CPN Tools and discuss the results of sonic simulation experiments.