11 Parameter Control in Evolutionary Optimisation
373
Fig. 11.8 A conceptual model of self-adaptive parameter control. The parameters are co-evolving
with the solutions of the problem
11.4.3 Self-Adapting Parameter Control
A search for good parameter values for an optimisation process is by itself an
optimisation problem. On the other hand, an optimisation algorithm that is being
tuned can be used to find good parameter values. A given optimisation algorithm
is designed to perform optimisation well and why not use it to find its own control
parameter values. Self-adjusting parameter control is a method where the search
for good control parameter values is performed simultaneously with the search of
optimal problem values, by encoding the control parameters into the chromosome
(as additional variables) (see Fig. 11.8).
In the case of EA, self-adjusting parameter control can be achieved by including
the control parameters in the chromosome alongside the decision variables. In this
way, an EA is able to try various control parameters and, through variation and
selection, choose control parameter values that are the most appropriate in a given
stage of optimisation. If some control parameter values lead the optimisation process
to candidate solutions with higher fitness, these control parameter values are more
likely to be inherited in the next generation. Control parameters that can be adjusted
in such a way include the type of evolutionary operators used and parameters
that determine their behaviour. For example, when using Gaussian mutation, the
chromosome can include the individual’s probability for mutation and the variance
of mutation [49].
One disadvantage of self-adjusting parameter control in EA is that the inclusion
of control parameters in the chromosome increases its size. This means that
chromosome modifications become computationally more expensive [50]. One way
to circumvent this issue is to use two populations. One population has individuals
with chromosomes consisting of decision variables, whilst the other population has
individuals with chromosomes consisting of control parameter values. This reduces
the complexity of the optimisation by splitting the search space into two smaller
search spaces.
Like the adaptive control, the self-adjusting control also uses feedback from the
optimisation and searches for parameter values that are optimal for the region of
373
Fig. 11.8 A conceptual model of self-adaptive parameter control. The parameters are co-evolving
with the solutions of the problem
11.4.3 Self-Adapting Parameter Control
A search for good parameter values for an optimisation process is by itself an
optimisation problem. On the other hand, an optimisation algorithm that is being
tuned can be used to find good parameter values. A given optimisation algorithm
is designed to perform optimisation well and why not use it to find its own control
parameter values. Self-adjusting parameter control is a method where the search
for good control parameter values is performed simultaneously with the search of
optimal problem values, by encoding the control parameters into the chromosome
(as additional variables) (see Fig. 11.8).
In the case of EA, self-adjusting parameter control can be achieved by including
the control parameters in the chromosome alongside the decision variables. In this
way, an EA is able to try various control parameters and, through variation and
selection, choose control parameter values that are the most appropriate in a given
stage of optimisation. If some control parameter values lead the optimisation process
to candidate solutions with higher fitness, these control parameter values are more
likely to be inherited in the next generation. Control parameters that can be adjusted
in such a way include the type of evolutionary operators used and parameters
that determine their behaviour. For example, when using Gaussian mutation, the
chromosome can include the individual’s probability for mutation and the variance
of mutation [49].
One disadvantage of self-adjusting parameter control in EA is that the inclusion
of control parameters in the chromosome increases its size. This means that
chromosome modifications become computationally more expensive [50]. One way
to circumvent this issue is to use two populations. One population has individuals
with chromosomes consisting of decision variables, whilst the other population has
individuals with chromosomes consisting of control parameter values. This reduces
the complexity of the optimisation by splitting the search space into two smaller
search spaces.
Like the adaptive control, the self-adjusting control also uses feedback from the
optimisation and searches for parameter values that are optimal for the region of