Abstraction and Training of Stochastic Graph Transformation Systems

Simulation of stochastic graph transformation systems (SGTS) allows us to analyse the model's behaviour. However, complexity of models limits our capability for analysis. In this paper, we aim to simplify models by abstraction while preserving relevant trends in their global behaviour. Based on a hierarchical graph model inspired by membrane systems, structural abstraction is achieved by "zooming out" of membranes, hiding their internal state. We use Bayesian networks representing dependencies on stochastic (input) parameters, as well as causal relationships between rules, for parameter learning and inference. We demonstrate and evaluate this process via two case studies, immunological response to a viral attack and reconfiguration in P2P networks.