11 Parameter Control in Evolutionary Optimisation
367
• irace—a software package with a number of automatic configuration procedures.
Its iterated procedures have been used to automatically configure various algorithms [31].
• SPOT—the sequential parameter optimisation package for R is a toolbox for
tuning and understanding simulation and optimisation algorithms [32]; see also
Chap. 10 for details.
• GGA—the gender-based genetic algorithm configurator, integrated with surrogate model for predicting high-performance regions in the parameter space [33].
• ParamILS—an automatic framework for the identification of performanceoptimising parameter settings [34].
• SMAC—the framework constructs explicit regression models to describe the
dependence of target algorithm performance on control parameter settings [35].
• Spearmint—software package to perform Bayesian optimisation which can be
used to find appropriate control parameter values in as few runs as possible [36].
The automated identification of control parameter values that these tools possess
recommends efficient values for static problems, where unchanged values of control
parameters are used through the whole optimisation process.
11.3.2 Parameter Control
The efficiency of optimisation with specific control parameter values depends on
the characteristics of the search space landscape. Some parameter values might be
good for separable functions, others for multimodal, convex, discontinuous and so
on. Usually, when population moves through the search space, such characteristics
change. For example, at the beginning of optimisation, the fitness landscape might
appear highly multimodal and irregular, but then the population converges towards
a region that contains a very narrow valley. In order for the algorithm to follow
this narrow valley efficiently, it needs to employ a different strategy than when
the landscape appeared multimodal. In this case different control parameter values
are appropriate for different stages of the optimisation process. This indicates that
no globally optimal control parameters exist and that one should use non-static
parameters that adapt during the search/optimisation process.
Moreover, different stages of the optimisation process require different control
parameter values [37]. It can be easily shown that exploration and exploitation
phases of the search process require different values of control parameters:
• In the beginning, in the exploration phase, large mutation rates and small
selective pressure are needed to make large jumps across the search space and
discover different areas of the search space.
• Towards the end, in the exploitation phase, small mutation rates and high selective pressure are required to focus the search around the promising discovered
regions of the search space.
367
• irace—a software package with a number of automatic configuration procedures.
Its iterated procedures have been used to automatically configure various algorithms [31].
• SPOT—the sequential parameter optimisation package for R is a toolbox for
tuning and understanding simulation and optimisation algorithms [32]; see also
Chap. 10 for details.
• GGA—the gender-based genetic algorithm configurator, integrated with surrogate model for predicting high-performance regions in the parameter space [33].
• ParamILS—an automatic framework for the identification of performanceoptimising parameter settings [34].
• SMAC—the framework constructs explicit regression models to describe the
dependence of target algorithm performance on control parameter settings [35].
• Spearmint—software package to perform Bayesian optimisation which can be
used to find appropriate control parameter values in as few runs as possible [36].
The automated identification of control parameter values that these tools possess
recommends efficient values for static problems, where unchanged values of control
parameters are used through the whole optimisation process.
11.3.2 Parameter Control
The efficiency of optimisation with specific control parameter values depends on
the characteristics of the search space landscape. Some parameter values might be
good for separable functions, others for multimodal, convex, discontinuous and so
on. Usually, when population moves through the search space, such characteristics
change. For example, at the beginning of optimisation, the fitness landscape might
appear highly multimodal and irregular, but then the population converges towards
a region that contains a very narrow valley. In order for the algorithm to follow
this narrow valley efficiently, it needs to employ a different strategy than when
the landscape appeared multimodal. In this case different control parameter values
are appropriate for different stages of the optimisation process. This indicates that
no globally optimal control parameters exist and that one should use non-static
parameters that adapt during the search/optimisation process.
Moreover, different stages of the optimisation process require different control
parameter values [37]. It can be easily shown that exploration and exploitation
phases of the search process require different values of control parameters:
• In the beginning, in the exploration phase, large mutation rates and small
selective pressure are needed to make large jumps across the search space and
discover different areas of the search space.
• Towards the end, in the exploitation phase, small mutation rates and high selective pressure are required to focus the search around the promising discovered
regions of the search space.