Skip to content. | Skip to navigation

Personal tools
Sections

Tobias Helms, Oliver Reinhardt, and Adelinde M Uhrmacher (2015)

Bayesian Changepoint Detection for Generic Adaptive Simulation Algorithms

In: SpringSim 2015 Annual Simulation Symposium.

Adaptive simulation algorithms are used to deal with changing computational demands of simulations due to state changes of the model and the environment. If such an algorithm is developed in a generic manner, i.e., it is not equipped by the developer with a function which decides how to switch its configuration, sophisticated techniques like machine  learning need to be exploited. Since adaptations can be costly, it is not practicable to adapt after each simulation step. Consequently, a fundamental challenge of generic adaptive simulation algorithms is to decide when to execute adaptations. For this, we present a dynamic algorithm based on Bayesian online changepoint detection. By observing  performance values regularly, this algorithm decides whether adaptations should be executed or not. We evaluate our approach based on a benchmark model defined in PDEVS and a model used in simulation studies defined in ML-Rules. Both modeling formalisms exhibit different dynamics and different requirements for adaptation and thus underline the generality of the adaptation strategy. Altogether, we present how the proposed Bayesian changepoint detection strategy helps balancing the effort required for adaptation, possible speed-up by this adaption, and the effectiveness of the machine learning algorithm.

Document Actions