11 Parameter Control in Evolutionary Optimisation
365
Fig. 11.4 Interdependence
between probabilities of
crossover and mutation (with
fixed number of generations
and population size). For
optimising the design of an
elliptic filter [18], it is shown
that the optimal tuned
(ensuring highest quality)
control parameters might be
μ = 10, n g = 100, p c = 0.7
and p m = 0.01
Table 11.1 Control
parameter settings as used in
[18]
Control parameter
Values
Population size
μ = 10
Number of generations n g = 100
Probability of crossover p c = 0.5, 0.6, 0.7, 0.8, 0.9
Probability of mutation p m = 0.01, 0.03, 0.05, 0.07, 0.09
11.3 Setting Approaches
The setting of the control parameters can be classified into two major approaches
[19], as presented in Fig. 11.5. The first one is parameter tuning, where the
parameters are tuned in advance of the actual optimisation process. The second one
is parameter control, where the parameters are modified online, during the optimisation process. In the latter case there are different strategies for the adaptation of
the control parameters.
Note that some newer works [13] suggest a slightly modified classification for
parameter control. Instead of making a distinction only amongst deterministic, adaptive and self-adaptive strategy, they suggest a classification into state-dependent,
success dependent, learning-dependent, self-adaptive and hyper-heuristics strategies. As currently most of the literature still relyies on the initial classification of
[19], this chapter follows that one, too.
For parameter tuning (see Sect. 11.3.1 for details) a large amount of literature
on finding good static control parameter values exists [20–23]. There are even
some frameworks to allow easy parameter tuning, a reference list of which is
being mentioned in the section below. The so-called parameter tuning mechanisms
involve checking several combinations and evaluating their efficiency. However,
optimal static parameter values for one problem may be much different for other
similarly looking problems. Namely, a small change in one parameter can cause
365
Fig. 11.4 Interdependence
between probabilities of
crossover and mutation (with
fixed number of generations
and population size). For
optimising the design of an
elliptic filter [18], it is shown
that the optimal tuned
(ensuring highest quality)
control parameters might be
μ = 10, n g = 100, p c = 0.7
and p m = 0.01
Table 11.1 Control
parameter settings as used in
[18]
Control parameter
Values
Population size
μ = 10
Number of generations n g = 100
Probability of crossover p c = 0.5, 0.6, 0.7, 0.8, 0.9
Probability of mutation p m = 0.01, 0.03, 0.05, 0.07, 0.09
11.3 Setting Approaches
The setting of the control parameters can be classified into two major approaches
[19], as presented in Fig. 11.5. The first one is parameter tuning, where the
parameters are tuned in advance of the actual optimisation process. The second one
is parameter control, where the parameters are modified online, during the optimisation process. In the latter case there are different strategies for the adaptation of
the control parameters.
Note that some newer works [13] suggest a slightly modified classification for
parameter control. Instead of making a distinction only amongst deterministic, adaptive and self-adaptive strategy, they suggest a classification into state-dependent,
success dependent, learning-dependent, self-adaptive and hyper-heuristics strategies. As currently most of the literature still relyies on the initial classification of
[19], this chapter follows that one, too.
For parameter tuning (see Sect. 11.3.1 for details) a large amount of literature
on finding good static control parameter values exists [20–23]. There are even
some frameworks to allow easy parameter tuning, a reference list of which is
being mentioned in the section below. The so-called parameter tuning mechanisms
involve checking several combinations and evaluating their efficiency. However,
optimal static parameter values for one problem may be much different for other
similarly looking problems. Namely, a small change in one parameter can cause