Karsten Wolf, Universität Rostock
Title: A fresh look into modular state spaces for Petri nets.
Abstract: Several Petri nets are modeled as a composition of smaller units, the modules. For such modular Petri nets, the construction of a modular state space has been proposed (Christensen and Petrucci, 2000). Modular state spaces differ from compositional state spaces in the order in which they compute the local behavior of the modules and a global data structure that records their interaction. As an effect, modular state spaces will never record local behavior that is not reached in the given composition.
In the beginning, the ability to distribute the computation of a modular state space was seen as a main advantage. In this talk, we shall point to other major advantages of the modular approach: replication and heterogeneity. If the composition contains several identical modules, only one local state space needs to be computed. This approach may lead to reduction even in systems where the symmetry method is not applicable.
If some of the modules enjoy special structural properties, we may switch to alternative (and more efficient) ways to represent the local state space even if other modules do not have such properties. We exercise this idea for acyclic modules and modules that happen to be Petri net state machines.
To make modular state space construction available to Petri nets without given modular structure, we reflect on ideas for cutting a plain Petri net into modules.
We believe that modular technique will turn out to be a useful complement to existing Petri net verification technology.

Sandjai Bhulai, Vrije Universiteit Amsterdam
Title: Performance evaluation in the age of learning and search
Abstract: Performance evaluation has traditionally started with models constructed by experts: stochastic models derived from assumptions, optimization problems formulated explicitly, and approximations designed to remain mathematically or computationally tractable. Increasingly, however, parts of this modeling process can themselves be learned from data or discovered through algorithmic search.
This shift raises questions that go beyond whether a machine-learning method can outperform a classical one. What do learned models capture that our traditional models miss? When is it useful to replace an explicitly constructed model or algorithm with a learned one? What happens to interpretability, robustness, guarantees, and our ability to reason about a system when we do so?
In this talk, I will explore these questions through examples from the modeling, prediction, optimization, and performance evaluation of communication and service systems. The examples illustrate different ways in which machine learning and computational search can extend the reach of classical approaches, but also ways in which they merely exchange analytical assumptions for empirical ones. Rather than arguing for analytical or learned models, the central question will be where the modeler should impose structure, where it can be discovered from data, and how the two can complement each other.