Simulator Adaption at Runtime for Component-Based Simulation Software

Helms, Tobias (2017) Simulator Adaption at Runtime for Component-Based Simulation Software. PhD thesis, Institut of Computer Science, University of Rostock.

Full text not available from this repository.
Official URL:


Component-based simulation software can provide many opportunities to compose and configure simulators, resulting in an algorithm selection problem for the user of this software. Further, as the state and structure of a model may vary during a simulation run, the computational demands might also change during runtime. Therefore, it is not only necessary to select a suitable simulator for executing a simulation run, but this selection must regularly be reconsidered to adapt the chosen simulator to changed computational demands. While this is a general and cross-cutting concern, most adaptation schemes for simulators are tailored to specific application scenarios that cannot be reused straightforwardly for other scenarios. Therefore, this thesis aims to automate the selection and adaptation of simulators at runtime in an application-independent manner. Further, it explores the potential of tailored and approximate simulators --- in this thesis concretely developed for the modeling language ML-Rules - supporting the effectiveness of the adaptation scheme. Specifically, for the automatic selection and adaptation of simulators at runtime, a flexible and generic adaptive simulator is developed and integrated into the modeling and simulation framework JAMES II. The adaptive simulator encapsulates available simulators applicable to a specific problem and employs reinforcement learning to explore and exploit the performance of these simulators. As it uses the encapsulated simulators to calculate the state transitions of a model, it is not restricted to any modeling language, but it can be applied to all modeling approaches available in JAMES II. To improve the learning efficiency of the adaptive simulator, state space generalization methods are applied. Further, different techniques to trigger adaptations are explored, e.g., a changepoint detection method monitoring the event throughput is integrated into the adaptive simulator. A pool of efficient simulators is a prerequisite for the effectiveness of the adaptive simulator. Therefore, in addition to the adaptive simulator itself, in this thesis tailored and approximate simulators are developed and explored concretely for the modeling language ML-Rules. Due to its expressiveness, it poses various computational challenges tackled by the developed simulators. The efficiency of these simulators is illustrated with complex ML-Rules models used in simulation studies.

Item Type: Thesis (PhD)
Projects: ESCeMMo