368
M. Antoniou et al.
The right choice of control parameter values depends on the problem and even the
problem instance that is solved and the algorithm that is employed for the search. So,
whenever a new optimisation problem or its new/different instance is introduced or
a new optimisation algorithm is employed, a new set of control parameter values is
needed. Surely, some preliminary experiments are often conducted to find (i.e. tune)
reasonable initial control parameter values. However, based on the implementation
of the adaptation, such tuning might not even be necessary since some types of
parameter control can find appropriate control parameter values in the early stage
of optimisation by themselves. The next section presents more details on different
adaptation strategies.
11.4 Parameter Control Strategies
When deciding on the proper parameter control strategy, several decisions have to
be made. First, one needs to decide what is the update trigger, which can be one of
the following:
• Number of fitness evaluations performed
• Time elapsed
• Progress in terms of absolute or relative fitness values
• Diversity measures
Next, it should be decided which strategy should be used to update the control
parameter values (their details are presented in the following sections):
• Deterministic (or time-dependent)
– It is usually time-dependent and the algorithm does not consider any feedback
from the optimisation process.
• Adaptive
– Here some statistical feedback from the optimisation process is used to
determine the direction and magnitude of change of the control parameter
value.
• Self-adaptive
– It treats control parameters as part of the optimisation (being a complex
optimisation problem) and EAs are used to find good values of the control
parameters (i.e. control parameters are encoded together with other variables
and are concurrently optimised).
M. Antoniou et al.
The right choice of control parameter values depends on the problem and even the
problem instance that is solved and the algorithm that is employed for the search. So,
whenever a new optimisation problem or its new/different instance is introduced or
a new optimisation algorithm is employed, a new set of control parameter values is
needed. Surely, some preliminary experiments are often conducted to find (i.e. tune)
reasonable initial control parameter values. However, based on the implementation
of the adaptation, such tuning might not even be necessary since some types of
parameter control can find appropriate control parameter values in the early stage
of optimisation by themselves. The next section presents more details on different
adaptation strategies.
11.4 Parameter Control Strategies
When deciding on the proper parameter control strategy, several decisions have to
be made. First, one needs to decide what is the update trigger, which can be one of
the following:
• Number of fitness evaluations performed
• Time elapsed
• Progress in terms of absolute or relative fitness values
• Diversity measures
Next, it should be decided which strategy should be used to update the control
parameter values (their details are presented in the following sections):
• Deterministic (or time-dependent)
– It is usually time-dependent and the algorithm does not consider any feedback
from the optimisation process.
• Adaptive
– Here some statistical feedback from the optimisation process is used to
determine the direction and magnitude of change of the control parameter
value.
• Self-adaptive
– It treats control parameters as part of the optimisation (being a complex
optimisation problem) and EAs are used to find good values of the control
parameters (i.e. control parameters are encoded together with other variables
and are concurrently optimised).