In Sect. 4, SPOT was run as an automatic tuner. Steps from the
automatic mode can be used in an interactive manner.
SPOT can be started with the command
spot (<configurationfile>, <task>)
where
configurationfile is the name of the SPOT configuration file (String) and
task is a STRING and can be one of the tasks init, seq, run, rep or auto.
SPOT can also be run in a meta mode to perform tuning over a set of problem
instances.
SPOT provides tools to perform
the following tasks:
- Initialize. An initial design is generated. This is usually the
first step during experimentation. The employed parameter region (ROI) and
the constant algorithm parameters (APD) have to be provided by the user.
SPOT's parameters are specified in the CONF file. Although it is
recommended to use the same basename for CONF, ROI, and APD files in order
to define a project, this is not mandatory. SPOT allows a flexible
combination of different filenames, e.g., one APD file can be used for
different projects.
- Run. This is
usually the second step. The optimization algorithm is started with
configurations of the generated design. Additionally information about the
algorithms problem design are used in this step. The algorithm writes
its results to the result file.
- Sequential step. A new
design, based on information from the result file, is
generated. A prediction model is used in this step. Several generic prediction models are available in SPOT
by default. To perform an efficient analysis, especially in
situations when only few algorithms runs are possible, user-specified prediction models can easily
be integrated into SPOT.
- Report. An analysis, based on information from the result file,
is generated. Since all data flow is stored in files, new report facilities can
be added very easily. SPOT contains some scripts to perform a basic
regression analysis and plots such as histograms, scatter plots, plots of the
residuals, etc.
- Automatic mode. In the automatic mode, the steps run
and sequential are performed after an initialization for a
predetermined number of times.
- Meta mode. In the meta mode, the tuning process is repeated for
several configurations. For example, tuning can be performed for different
starting points
, several dimensions, or randomly chosen problem
instances.
Subsections
bartz
2010-07-08