378
M. Antoniou et al.
• Multi-population approaches, where many small populations track their own
peaks as the environment changes
The above strategies require a parameter control approach.
11.5.3 Optimisation Under Uncertainty
The presence of a range of uncertainties has to be taken into account for solving
many real-world applications with EAs. Jin and Branke [73] categorise the uncertainties that influence EA performance into four types:
• When noise occurs in the fitness function
• In case of design and environmental parameters subject to changes after the
optimisation
• When fitness function is an approximation
• When the optimum changes over time (as in dynamic optimisation, see
Sect. 11.5.2)
Methodologies to addressing noisy fitness function are explicit averaging by
calculating the average of the fitness values over a number of randomly sampled disturbances (see, e.g., [74, 75]), implicit averaging sample size as an inverse function
of the population size [76], fitness inheritance where the offspring inherits also the
mean and standard deviation of the objective value [77] and selection modification
[78]. These methods assume that the search space follows a homogeneous noise
distribution such as a uniform or a normal distribution [79].
11.5.4 Multi-objective Optimisation
Multi-objective and also many-objective optimisation (see details in Chap. 8)
approaches are used for optimisation problems where several criteria need to be
optimised that are equally treated and not merged (e.g. by weights) into one single
objective. The output of multi-objective optimisation is a set of solutions that
approximates the Pareto front. The main difference here is that there is no unique
measure that would indicate how good a current approximation of the Pareto front is.
Unlike in single-objective case where fitness can be used to measure this. Therefore,
in multi-objective case, adaptive parameter control is a bit more complicated to
design and additional considerations are needed to design the phenotype feedback
collection part.
One possibility for assessing in what stage the optimisation process is, is to
monitor the proportion of non-dominated solutions in the population [80], or
convergence detection [81]. This quantity typically increases during optimisation
process and can be used to guide the parameter control. The most common
M. Antoniou et al.
• Multi-population approaches, where many small populations track their own
peaks as the environment changes
The above strategies require a parameter control approach.
11.5.3 Optimisation Under Uncertainty
The presence of a range of uncertainties has to be taken into account for solving
many real-world applications with EAs. Jin and Branke [73] categorise the uncertainties that influence EA performance into four types:
• When noise occurs in the fitness function
• In case of design and environmental parameters subject to changes after the
optimisation
• When fitness function is an approximation
• When the optimum changes over time (as in dynamic optimisation, see
Sect. 11.5.2)
Methodologies to addressing noisy fitness function are explicit averaging by
calculating the average of the fitness values over a number of randomly sampled disturbances (see, e.g., [74, 75]), implicit averaging sample size as an inverse function
of the population size [76], fitness inheritance where the offspring inherits also the
mean and standard deviation of the objective value [77] and selection modification
[78]. These methods assume that the search space follows a homogeneous noise
distribution such as a uniform or a normal distribution [79].
11.5.4 Multi-objective Optimisation
Multi-objective and also many-objective optimisation (see details in Chap. 8)
approaches are used for optimisation problems where several criteria need to be
optimised that are equally treated and not merged (e.g. by weights) into one single
objective. The output of multi-objective optimisation is a set of solutions that
approximates the Pareto front. The main difference here is that there is no unique
measure that would indicate how good a current approximation of the Pareto front is.
Unlike in single-objective case where fitness can be used to measure this. Therefore,
in multi-objective case, adaptive parameter control is a bit more complicated to
design and additional considerations are needed to design the phenotype feedback
collection part.
One possibility for assessing in what stage the optimisation process is, is to
monitor the proportion of non-dominated solutions in the population [80], or
convergence detection [81]. This quantity typically increases during optimisation
process and can be used to guide the parameter control. The most common