Information Technology Reference
In-Depth Information
regions in the parameter space which lead to a certain output behavior. Their approach
is based on support vector machines (SVM) and active learning, i.e., they aim at an
intelligent selection of new points in the parameter space in order to maximize “the
amount of new information obtained” [2, p. 83]. As applications they use asteroid col-
lision simulation and simulation of the Earth's magnetosphere. They report an increase
of the efficiency over standard gridding (
2
).
Hoad et al. [4] introduce an algorithm for the automated selection of the number
of replications for discrete-event simulation in order to achieve a certain accuracy for
simulation output measures taking into account confidence intervals. They apply the
approach to different statistical distributions and to a set of simulation models. The
authors report that the algorithm is effective in selecting the needed number replications
in order to cover the expected mean at a given level of precision.
Similar to some of the related approaches, we apply machine learning in combination
with simulation. In this work, machine learning is not used to discover knowledge from
simulation results but to learn a classifier for the estimation of statistical properties. In
our approach, we take into account statistical tests and the development of their results
for the control of simulation runs.
×
to
6
×
3
Control of Simulation Experiments
In this section, we briefly describe the project context of the approach presented in this
paper. The goal of the associated project
AssistSim
is the provision of support function-
alities for the performance of simulation studies. Assistance is intended for planning,
execution, and analysis of simulation studies. The first aspect - planning assistance -
aims at capturing relevant information for a simulation study, e.g., identification of the
objects of investigation including parameters as well as their domains, and selection of
measurements and target functions. Details about this aspect are planned to be published
in a separate paper by our project partners.
The aim of the second aspect - the execution assistance - is the automated opera-
tion and control of the simulation system, i.e., the automated execution of simulation
runs. This phase is partially connected with the analysis assistance as simulation con-
trol depends on intermediate results of simulation runs. However, in the current project,
we restrict the analysis assistance to a relevant set of functions for simulation control.
A thoroughly designed analysis assistance for the investigation of a large result set of
simulation studies is planned to be part of a follow-up project.
The essential task of the simulation execution assistance is the systematic execution
of the different settings of the planned experiments. It is distinguished between three
different kinds of simulation studies:
1. Exploration: The parameter space has to be explored and interesting findings should
be captured.
2. Optimization: Parameter configurations which are expected to lead to good results
w.r.t. a target functions should be identified.
3. Comparison: Two or more parameter configurations of a simulation model (or dif-
ferent simulation models) should be compared identifying the best one or ranking
the variants w.r.t. a target function.
