11 Parameter Control in Evolutionary Optimisation
361
algorithm control parameter values depend on the current state of the optimisation
process and thus change over time [11, 12].
As the adaptation of the algorithm control parameters depends on different
scenarios, one should consider their influence (further details are available in
Sect. 11.4): deterministic (time-dependent, feedback-free), self-adaptive (encoding
control parameters with other variables) and adaptive (using statistical feedback
from the optimisation process). The automatic setup of algorithms and their control
parameters is one of the prerequisites to allow ease of use of the complex industrial
optimisation tools.
One should differentiate between control parameter (Sect. 11.2), which denotes
the parameter of the algorithm that drives the algorithm’s behaviour, and the
parameter control (Sect. 11.3), which denotes the online process of changing the
control parameter(s). The main goals of parameter control are the identification of
good control parameter values on-the-fly and tracking of good control parameter
values as they change during the optimisation process [13].
11.2 Control Parameters
This section presents typical control parameters that can be found in evolutionary
algorithms, along with their characteristics. This is followed by a description of their
influence on the algorithm’s performance and also the interdependence amongst
different control parameters is presented.
11.2.1 Typical Control Parameters
There are obviously different control parameters [14], with different effects, for different optimisation algorithms. As an example, for the most typical implementations
of EAs, one can adapt:
• Population size
– Also referred as μ.
– The number of individuals (solutions) that are considered (and evaluated) in
parallel and are combined to form new solutions.
– In general, larger population size gives larger variability within the population,
but increases the computational time and slows down the convergence (when
the number of generations remains the same).
• Offspring size
– Also referred as λ.
– The number of offspring (resulting solutions) that are produced at each
generation.
361
algorithm control parameter values depend on the current state of the optimisation
process and thus change over time [11, 12].
As the adaptation of the algorithm control parameters depends on different
scenarios, one should consider their influence (further details are available in
Sect. 11.4): deterministic (time-dependent, feedback-free), self-adaptive (encoding
control parameters with other variables) and adaptive (using statistical feedback
from the optimisation process). The automatic setup of algorithms and their control
parameters is one of the prerequisites to allow ease of use of the complex industrial
optimisation tools.
One should differentiate between control parameter (Sect. 11.2), which denotes
the parameter of the algorithm that drives the algorithm’s behaviour, and the
parameter control (Sect. 11.3), which denotes the online process of changing the
control parameter(s). The main goals of parameter control are the identification of
good control parameter values on-the-fly and tracking of good control parameter
values as they change during the optimisation process [13].
11.2 Control Parameters
This section presents typical control parameters that can be found in evolutionary
algorithms, along with their characteristics. This is followed by a description of their
influence on the algorithm’s performance and also the interdependence amongst
different control parameters is presented.
11.2.1 Typical Control Parameters
There are obviously different control parameters [14], with different effects, for different optimisation algorithms. As an example, for the most typical implementations
of EAs, one can adapt:
• Population size
– Also referred as μ.
– The number of individuals (solutions) that are considered (and evaluated) in
parallel and are combined to form new solutions.
– In general, larger population size gives larger variability within the population,
but increases the computational time and slows down the convergence (when
the number of generations remains the same).
• Offspring size
– Also referred as λ.
– The number of offspring (resulting solutions) that are produced at each
generation.