A generic adaptive simulation algorithm for component-based simulation systems

Abstract

The state of a model may strongly vary during simulation, and with it also the simulation's computational demands. Adapting the simulation algorithm to these demands at runtime can therefore improve the overall performance. Although this is a general and cross-cutting concern, only few simulation systems offer re-usable support for this kind of runtime adaptation. We present a flexible and generic mechanism for the runtime adaptation of component-based simulation algorithms. It encapsulates simulation algorithms applicable to a given problem and employs reinforcement learning to explore the algorithms' suitability during a simulation run. We evaluate the approach by executing models from two modeling formalisms used in computational biology.

Syntax and Semantics of a Multi-Level Modeling Language

“…However, this also facilitates the use of alternative kinetics, e.g., MichaelisMenten or Hill kinetics [24]. (13) denotes that patterns composed with the + operator have to be matched together as well. Also, bindings from one pattern instantiation are made visible in other pattern instantiations by using the same σ.…”

“…Furthermore, we added multi-key data structures and different hash algorithms to efficiently access and compare species with specific properties. We also applied an adaptive simulation algorithm on top of the ML-Rules simulator to change its configuration during runtime to improve its performance [13]. Besides these implementation optimizations, we developed approximate variants of the simulation algorithm based on τ -leaping [14] and hybrid algorithms.…”

Automatic Runtime Adaptation for Component-Based Simulation Algorithms

“…For preprocessing, we use a static state space generalization with configuration c 2 . Here, in contrast to Helms et al [2013], the 36 actions correspond to 36 configurations of the ML-Rules SA, combined with only one adaptation condition (adapt every 1,000 events). In Helms et al [2013], we used 12 SA configurations and three adaptation conditions.…”

“…Here, in contrast to Helms et al [2013], the 36 actions correspond to 36 configurations of the ML-Rules SA, combined with only one adaptation condition (adapt every 1,000 events). In Helms et al [2013], we used 12 SA configurations and three adaptation conditions. Our previous results show that, at least in this setting, considering multiple step count adaptation conditions only worsened the learning efficiency and did not improve the overall performance.…”

HPC simulations of information propagation over complex networks

“…The runtime adaptation mechanism has also been used in simulations, including continuous simulations [103] and parallel discrete event simulations [40]. Uhrmacher's group has proposed a more general concept of adaptive simulator to allow the adaptive mechanism to be applied across simulation domains [62]. "Plug-in Type" interfaces provide a description of simulation parameters and simulation components (i.e.…”