UNIT 3 FITNESS, SELECTION,

AND POPULATION
MANAGEMENT
Ami Munshi
Syllabus
Course Outcomes
References
Fitness, Selection, and Population
Management
 There are two main forces that form the basis of
evolutionary systems:
 Variation operators (recombination and mutation)
create the necessary diversity within the population,
and thereby facilitate novelty.
 Selection acts as a force increasing the mean quality of
solutions in the population
Population model
Population Management Model
 We have focused on
 the way that potential solutions are represented to give a population of diverse
individuals,
 the way that variation (recombination and mutation) operators work on those
individuals to yield offspring
 These offspring will generally inherit some of their parents’ properties but also
differ slightly from them, providing new potential solutions to be evaluated
 Now we turn our attention to the second important element of the
evolutionary process –
 the differential survival of individuals to compete for resources and take part in
reproduction, based on their relative fitness
Population Management Models
 Two different models of population management
are found in the literature:
 the generational model
 the steady-state model
Generational model
 In each generation we begin with a population of size μ,
 From this population, a pool of parents is selected
 Every member of the pool is a copy of something in the population
 But the proportions will probably differ, with (usually) more copies
of the ‘better’ parents
 Next, λ offspring are created from the pool by the application of
variation operators, and evaluated
 After each generation, the whole population is replaced by μ
individuals selected from its offspring, which is called the next
generation
Steady state model
 In the steady-state model, the entire population is not changed at
once, but rather a part of it
 In this case, old individuals are replaced by new ones, the offspring
 The proportion of the population that is replaced is called the
generational gap, and is equal to λ/μ
 The operators that are responsible for this competitive element of
population management work on the basis of an individual’s fitness
 As a direct consequence, these selection and replacement operators
work independently of the problem representation chosen
 Generational Loop in Evolutionary
Agorithm

Ref: SpringerBookChapter Population Management.pdf

Generational Loop in Evolutionary
Agorithm
 One begins by generating a set of μ solutions to the problem,
usually in a purely random fashion
 This set of solutions is called a population by analogy with a group
of living being
 In the same way, a solution to the problem is an individual
 Evolutionary algorithms repeat the next loop (called a generational
loop) until a stopping criterion is met
 This is either set in advance, for example, the number of times the
generational loop is repeated, or decided on the basis of the
diversity of individuals present in the population
Generational Loop in Evolutionary
Algorithm
 First, a number of solutions from the population are selected to be used for
breeding
 This is achieved by a selection operator for reproduction
 The purpose of this operator is to favour the individuals that are well adapted to
their environment (those with the best fitness function) at the expense of those that
are weaker, sick, and ill-adapted similar to what happens in nature
 The selected individuals are then mixed together (e.g., in pairs) using a crossover
operator to form λ new solutions called offspring
 They undergo random modifications by means of a mutation operator
 Finally, the new solutions are evaluated, and a selection operator for survival
eliminates λ solutions from the μ + λ available to reduce to a new population of μ
individuals
Parent Selection
Parent Selection/Selection for
reproduction
 Fitness Proportional Selection (FPS)/Cannonical
Selection
 Rank based Selection
 Roulette Wheel Selection
 Stochastic Uniform Sampling
 Tournament Selection
 Uniform Parent Selection
Fitness Proportionate Selection
 Probability that an individual 𝑖 is selected for mating
depends on its absolute fitness value compared to the
absolute fitness values of the rest of the population
 Observing that the sum of the probabilities over the whole
population must equal 1 the selection probability of
individual 𝑖 using FPS is
Fitness Proportionate Selection
 However, it has been recognised that there are some problems with
this selection mechanism
 Outstanding individuals take over the entire population very quickly
 This makes it less likely that the algorithm will thoroughly search the space of
possible solution
 This phenomenon is often observed in early generations, when many of the
randomly created individuals will have low fitness, and is known as premature
convergence
 When fitness values are all very close together, there is almost no
selection pressure, so selection is almost uniformly random, and having a
slightly better fitness is not very ‘useful’ to an individual
 The mechanism behaves differently if the fitness function is transposed
 This is shown in the table below
 Fitness Proportionate Selection
 This last point above is illustrated in Table below,
 which shows three individuals and a fitness function with f(A)=1, f(B) = 4, and f(C) = 5
 Transposing this fitness function changes the selection probabilities, while the shape of the
fitness landscape, and hence the location of the optimum, remains the same

Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
 Evolutionary cycle by hand- Example
𝑓 𝑖
𝑝𝑖 = , 𝑤ℎ𝑒𝑟𝑒 𝑃 = 𝑃𝑜𝑝𝑢𝑙𝑎𝑡𝑖𝑜𝑛
σ𝑗∈𝑃 𝑓 𝑗

Genotype

Phenotype

Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
Example-DIY
1. You are given the fitness function f(x) = x2 +10 and a population
of three individuals {a,b,c}. When decoded their genes when
decoded give the values 1, 2 and 3 respectively. When you pick a
single parent using Fitness Proportionate Selection, what is the
probability that it is b?
2. Calculate the probabilities of selecting b via fitness proportionate
selection if the new fitness function is f'(x) = x2
3. What is the probability of selecting b via binary tournament
selection for the example above? How does it change when the
fitness function has it’s values reduced by 10?
Ranking Selection
 Rank-based selection is another method that was inspired by the
observed drawbacks of fitness proportionate selection
 It preserves a constant selection pressure by sorting the population
on the basis of fitness
 Then allocating selection probabilities to individuals according to
their rank, rather than according to their actual fitness values
 Let us assume that the ranks are numbered so
 the best has rank μ-1 and the worst has rank 0
 The mapping from rank number to selection probability can be done
in many ways, for example, linearly or exponentially decreasing
Ranking Selection
 The usual formula for calculating the selection
probability for linear ranking schemes is
parameterised by a value s (1 < s ≤ 2)
 Since the best individual has rank μ − 1, and the
worst has rank 0, then the selection probability for
an individual of rank 𝑖 is
Ranking Selection

 Note that
 the first term will be constant for all individuals (it is there to
ensure the probabilities add to one)
 Since the second term will be zero for the worst individual
(with rank 𝑖 = 0), it can be thought of as the ‘baseline’
probability of selecting that individual
 Ranking Selection
 In Table below an example of how the selection probabilities differ for
a population of 𝜇 = 3 different individuals with fitness proportionate
selection and rank-based selection with different values of 𝑠
 Ranking Selection
 When the mapping from rank to selection probabilities is linear, only limited selection
pressure can be applied
 This arises from the assumption that, on average, an individual of median fitness should
have one chance to be reproduced,
 which in turn imposes a maximum value of s = 2.
 (Since the scaling is linear, letting s > 2 would require the worst to have a negative selection
probability if the probabilities are to sum to unity)
 If a higher selection pressure is required, i.e., more emphasis on individuals of above-
average fitness, an exponential ranking scheme is often used, of the form

The normalisation factor c is chosen so that the sum of the probabilities is unity, i.e., it is a
function of the population size
Implementing the selection probabilities
 The description above provides two alternative schemes
for deciding a probability distribution that defines the
likelihood of each individual in the population being
selected for reproduction
 The mating pool of parents is sampled from the
selection probability distribution
 The simplest way of achieving this sampling is known as
the roulette wheel algorithm
Roulette Wheel Selection
 Conceptually this is the same as repeatedly spinning a one-armed roulette
wheel, where the sizes of the holes reflect the selection probabilities
 In general, the algorithm can be applied to select λ members from the set
of μ parents into a mating pool
 To illustrate the workings of this algorithm,
 we will assume some order over the population (ranking or random) from 1 to μ,
 so that we can calculate the cumulative probability distribution, which is a list of
values [a1, a2,...,aµ] such that
 𝑎𝑖 = σ𝑖𝑖 𝑃𝑠𝑒𝑙 (𝑖),
 where 𝑃𝑠𝑒𝑙 (𝑖) is defined by the selection distribution — fitness proportionate or
ranking
 Note that this implies 𝑎µ = 1
 Roulette Wheel Selection
 Consider a wheel and divide it into m divisions
Where m is the number of chromosomes in populations
 Area occupied by each chromosome is proportional to its fitness
value
Survival Percentage
Survival Points
Points
Chromosome 1 28 28.95
29%
35%
Chromosome 2 23 23.7%
Chromosome 3 12 12.4%
Chromosome 4 34 35.1% 12% 24%

Total 97
Chromosome 1 Chromosome 2 Chromosome 3 Chromosome 4
 Roulette Wheel Selection
 Circular wheel is divided into pies
 A fixed point is chosen on the wheel
circumference
 Wheel is rotated
 The region of the wheel which
comes in front of the fixed point is
chosen as the parent
 For the second parent, the same
process is repeated
 It is to be noted that fitness
proportionate selection methods
don’t work for cases where the
fitness can take a negative value

Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
 Roulette Wheel Selection
 Fitter individual has a greater pie on the wheel and therefore a greater chance of landing in
front of the fixed point when the wheel is rotated
 Therefore, the probability of choosing an individual depends directly on its fitness
 Algorithm:
 Let there be 𝑚 chromosomes
 Find total of all the fitness values, say sum
 Find individual probability for each chromosome 𝑚𝑖 as 𝑃𝑖 = 𝑓𝑖 /𝑠𝑢𝑚
 Calculate cdf for mi as = σ 𝑝𝑖
 Generate a random number , say 𝑟, between 0 𝑎𝑛𝑑 1
 If 𝑟 < 𝐶1
 Then select 𝑚1
 Else if 𝐶𝑖−1 < 𝑟 < 𝐶𝑖 :
 Then select 𝑚𝑖
Example on Roulette Wheel Selection
 Let there be 𝑚 chromosomes
Chromosome Fitness Probabilit CDF  Find total of all the fitness values, say 𝑠𝑢𝑚
(𝑚𝒊 ) value y (𝑷𝒊 = (𝑪𝒊 )  Find individual probability for each chromosome
(𝐹𝒊 ) 𝑭𝒊 / σ 𝐹𝑖 ) mi as 𝑃𝑖 = 𝑓𝑖 /𝑠𝑢𝑚
𝑖=1 A 12 0.12 0.12  Calculate cdf for 𝑚𝑖 as 𝑃𝑖 = 𝑓𝑖 /𝑠𝑢𝑚
 Generate a random number , say 𝑟, between
𝑖=2 B 20 0.20 0.32 0 𝑎𝑛𝑑 1
 If 𝑟 < 𝐶1
𝑖=3 C 25 0.25 0.57  Then select 𝑚1
 Else if 𝐶𝑖−1 < 𝑟 < 𝐶𝑖 :
𝑖=4 D 10 0.10 0.67  Then select 𝑚𝑖
 Example:
 Say 𝑟 = 0.11 is generated. Then chromosome
𝑖=5 E 33 0.33 1 𝑚1 will be selected. That is chromosome A is
selected
 If 𝑟 = 0.62 is generated. It lies between
෍ 𝐹𝑖 𝑚3 𝑎𝑛𝑑 𝑚4 . In this case, chromosome 𝑚4 which is
chromosome D is chosen
= 100
 Stochastic Universal Sampling (SUS)
 Similar to Roulette wheel selection
 Instead of having just one fixed point, we have multiple fixed points
 Therefore, all the parents are chosen in just one spin of the wheel
 Encourages the highly fit individuals to be chosen at least once
 It is to be noted that fitness proportionate selection methods don’t work for cases where the fitness can take a negative
value

Ref: https://www.tutorialspoint.com/genetic_algorithms/genetic_algorithms_parent_selection.htm
Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
Tournament Selection
 The previous two selection methods and the algorithms used to sample from their probability distributions
relied on a knowledge of the entire population
 However, in certain situations,
 if the population size is very large, or
 if the population is distributed in some way (perhaps on a parallel system)
 obtaining this knowledge is either highly time consuming or at worst impossible
 Tournament selection is an operator with the useful property that
 it does not require any global knowledge of the population
 nor a quantifiable measure of quality
 Instead it only relies on an ordering relation that can compare and rank any two individuals
 It is therefore conceptually simple and fast to implement and apply
 The application of tournament selection to select λ members of a pool of μ individuals works according to
the procedure shown in the algorithm below
Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
 Tournament Selection
 In K-Way tournament selection select K individuals from the population at random
 Select the best out of these to become a parent
 The same process is repeated for selecting the next parent
 It can even work with negative fitness values

Ref: https://www.tutorialspoint.com/genetic_algorithms/genetic_algorithms_parent_selection.htm
 Rank Selection
 The usual formula for calculating the selection probability for linear
ranking schemes is parameterised by a value s (1 < s ≤ 2)
 In the case of a generational EA, where μ = λ, this can be interpreted
as the expected number of offspring allotted to the fittest individual
 Since this individual has rank μ − 1, and the worst has rank 0, then the
selection probability for an individual of rank i is
 Uniform Parent Selection
 In some dialects of EC it is common to use mechanisms such that each individual has
the same chance to be selected
 In Evolutionary Programming,
 usually there is no recombination
 only mutation
 and parent selection is deterministic
 In particular, each parent produces exactly one child by mutation
 Evolution Strategies are also usually implemented with uniform random selection of
parents into the mating pool

Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
Over-selection for Large Population
 In some cases it may be desirable to work with extremely large populations
 Sometimes this could be for technical reasons – for example,
 there has been a lot of interest in implementing EAs using graphics cards (GPUs),
which offer similar speed-up to clusters or supercomputers, but at much lower
cost
 However, achieving the maximum potential speed-up typically depends on
having a large population on each processing node
 If the potential search space is enormous it might be a good idea to use a
large population to avoid ‘missing’ promising regions in the initial random
generation, and thereafter to maintain the diversity needed to support
exploration
Over-selection for Large Population
 In some cases it may be desirable to work with extremely large populations
 Sometimes this could be for technical reasons – for example,
 there has been a lot of interest in implementing EAs using graphics cards (GPUs), which
offer similar speed-up to clusters or supercomputers, but at much lower cost
 However, achieving the maximum potential speed-up typically depends on having a
large population on each processing node
 If the potential search space is enormous it might be a good idea to use a large
population to avoid ‘missing’ promising regions in the initial random generation, and
thereafter to maintain the diversity needed to support exploration
 Often a method called over-selection is used for population sizes of 1000 and
above
Over-selection for Large Population
 In this method, the population is first ranked by
fitness and then divided into two groups,
 the top x% in one and the remaining (100−x)% in the
other
 When parents are selected,
 80% of the selection operations choose from the first
group, and
 the other 20% from the second
 Over-selection for Large Population
 Author Koza provides rule of thumb values for x depending on
the population size as shown in Table below
 As can be seen, the number of individuals from which the majority
of parents are chosen stays constant, i.e., the selection pressure
increases dramatically for larger populations

Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
Survivor Selection
Survivor Selection
 The last key operator in genetic algorithms is selection for survival
 which aims to bring the population back to its initial size of μ individuals, after λ new solutions have been generated
 Several selection policies have been devised, depending on the values chosen for the parameters μ and λ
 The survivor selection mechanism is
 responsible for managing the process of reducing the working memory of the EA from a set of μ parents and λ offspring
to a set of μ individuals forming the next generation
 In principle, any of the mechanisms introduced for parent selection could be also used for selecting survivors
 This step in the main evolutionary cycle is also called replacement
 Replacement strategies can be categorized according to
 Fitness based replacement
 Age based replacement
 Generational replacement
Generational replacement
 The simplest policy for selecting the individuals who will survive is to
 generate the same number of offspring as there are individuals in the
population (λ = μ)
 The population at the beginning of the new generational loop is
 made up only of the offspring
 the initial population disappears
 With such a choice, it is necessary to have a selection operator for
reproduction that favors the best solutions
 This means the best individuals are able to participate in the creation of
several offspring
 While some of the worst are excluded from the reproduction process
Evolutionary Strategy
 The evolutionary strategy (μ, λ) consists in generating
numerous offspring (λ>μ) and in only keeping the μ best
offspring for the next generation
 The population is therefore completely changed from one
iteration of the generational loop to the next
 This strategy leads to a bias in the choice of the fittest
individuals from one generation to the next
 So, it is compatible with a uniform selection operator for
reproduction.
Age based replacement
 The basis of these schemes is that the fitness of individuals is not taken into account during the
selection
 Instead, they are designed so that each individual exists in the population for the same
number of EA iterations
 Note
 Since fitness is not taken into account, the mean, and even best fitness of any given generation, may be
lower than that of its predecessor
 While slightly counterintuitive, this is not a problem as long as
 it does not happen too often,
 and may even be beneficial if the population is concentrated around a local optimum
 A net increase in the mean fitness over time therefore relies on
 having sufficient selection pressure when selecting parents into the mating pool
 using variation operators that are not too disruptive
 Age-based replacement is the strategy used in the simple Genetic Algorithm
Age based replacement
 Since the number of offspring produced is the same
as the number of parents (μ = λ), each individual
exists for just one cycle, and the parents are simply
discarded, to be replaced by the entire set of
offspring
 In this case the strategy takes the form of a first-in-
first-out (FIFO) queue
Fitness based replacement
 A wide number of strategies based on fitness have been proposed for
choosing which
 μ of the μ parents and λ offspring should go forward to the next generation
 Some also take age into account
 Some methods
 Replace worst (GENITOR)
 Elitism
 Round-robin tournament
 (μ + λ) Selection
 (μ, λ) Selection
Replace worst (Genitor)
 In this scheme the worst λ members of the population are
selected for replacement
 This can lead to very rapid improvements in the mean
population fitness
 However it can also lead to premature convergence as the
population tends to rapidly focus on the fittest member
currently present
 For this reason it is commonly used in conjunction with large
populations and/or a “no duplicates” policy
Elitism
 This scheme is commonly used in conjunction with age-based
and stochastic fitness-based replacement schemes
 This is done to prevent the loss of the current fittest member
of the population
 In essence a trace is kept of the current fittest member, and
it is always kept in the population
 Thus if it is chosen in the group to be replaced, and none of
the offspring being inserted into the population has equal or
better fitness than the current fittest member, then it is kept
and one of the offspring is discarded
Round robin tournament
 The method works by holding pairwise tournament competitions in
round-robin format
 Each individual is evaluated against q others randomly chosen from
the merged parent and offspring populations
 For each comparison, a “win” is assigned if the individual is better
than its opponent
 After finishing all tournaments, the μ individuals with the greatest
number of wins are selected
 Typically, q = 10 is recommended in Evolutionary Programming
 (μ + λ) Selection
 The name and the notation of the (μ + λ) selection comes from Evolution
Strategies
 In general, it refers to the case where the set of offspring and parents are
merged
 They are ranked according to (estimated) fitness
 Then the top μ are kept to form the next generation
 This strategy can be seen as a generalisation of
 GENITOR method (μ>λ) and
 Round-Robin tournament (μ = λ)
 In Evolution Strategies λ>μ
 A great offspring surplus (typically λ/μ ≈ 5 to 7) Induces a large selection
pressure
(μ, λ) Selection
 The (μ, λ) strategy used in Evolution Strategies where typically λ>μ children are
created from a population of μ parents
 This method works on a mixture of age and fitness
 The age component means that all the parents are discarded, so no individual is
kept for more than one generation
 The fitness component comes from the fact that the λ offspring are ranked according
to the fitness, and the best μ form the next generation
 In Evolution Strategies, (μ, λ) selection, is generally preferred over (μ + λ) selection
for the following reasons
 The (μ, λ) discards all parents and is therefore in principle able to leave (small) local
optima
 This may be advantageous in a multimodal search space with many local optima
Comparison of (μ, λ) with (μ+λ)
 In Evolution Strategies, (μ, λ) selection, is generally
preferred over (μ + λ) selection for the following reasons
 (μ, λ) selection discards all parents and is therefore in principle
able to leave (small) local optima
 This may be advantageous in a multimodal search space with many
local optima
 If the fitness function is not fixed, but changes in time, the (μ+λ)
selection preserves outdated solutions, so it is not able to follow
the moving optimum well
 (μ + λ) selection hinders the self-adaptation mechanism used to
adapt strategy parameters
Selection Pressure
 Notion of selection pressure
 using an intuitive description that
 as selection pressure increases,
 so fitter solutions are more likely to survive, or be chosen as parents,
 and less-fit solutions are correspondingly less likely
 Selection pressure is defined as
 the degree to which the better individuals are favored
 Higher the selection pressured, the more the better individuals are favored
 This selection pressure drives the GA to improve the population fitness over the successive
generations
 Convergence rate of GA is largely determined by the magnitude of the selection
pressure, with higher selection pressures resulting in higher convergence rates
Selection pressure
 Genetic Algorithms should be able to identify optimal or nearly
optimal solutions under a wide range of selection scheme pressure
 If the selection pressure is too low, the convergence rate will be slow, and
the GA will take unnecessarily longer time to find the optimal solution
 If the selection pressure is too high, there is an increased change of the
GA prematurely converging to an incorrect (sub-optimal) solution
 In addition to providing selection pressure
 selection schemes should also preserve population diversity, as this helps
to avoid premature convergence
Multimodal problems
 Adaptive Surface
 The adaptive-landscape model is a useful
way of combining
 Evolution
 population genetics
 environmental selective pressures into a single
view of speciation
 On this landscape the height dimension
belongs to fitness:
 high altitude stands for high fitness
 The other two (or more, in the general case)
dimensions correspond to biological traits

Ref: https://www.zoology.ubc.ca/~bio310/121T_files/06S_adaptivelandscape.pdf
Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
 Adaptive Surface
 xy-plane holds all possible trait
combinations
 z values show their fitnesses
 Hence, each peak represents a range of
successful trait combinations, while
troughs belong to less fit combinations
 A given population can be plotted as a
set of points on this landscape, where
each dot is one individual realising a
possible trait combination.

Ref: https://www.zoology.ubc.ca/~bio310/121T_files/06S_adaptivelandscape.pdf
 Adaptive Surface
 Evolution is then the process of
gradual advances of the
population to high-altitude
areas, powered by variation and
natural selection
 Our familiarity with the physical
landscape on which we exist
naturally leads us to the concept
of multimodal problems
Ref: https://www.zoology.ubc.ca/~bio310/121T_files/06S_adaptivelandscape.pdf
 Exploration and Exploitation
 To illustrate how an EA typically works, we will assume a onedimensional objective function to be maximized
 Figure below shows three stages of the evolutionary search, showing how the individuals might typically be distributed in the
beginning, somewhere halfway, and at the end of the evolution
 In the first stage directly after initialisation, the individuals are randomly spread over the whole search space
 After only a few generations this distribution changes: because of selection and variation operators the population abandons low-
fitness regions and starts to climb the hills
 Yet later (close to the end of the search, if the termination condition is set appropriately), the whole population is concentrated around
a few peaks, some of which may be suboptimal
 In principle it is possible that the population might climb the wrong hill, leaving all of the individuals positioned around a local but not
global optimum

Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015
Exploration and Exploitation
 Concept of exploration and exploitation, these notions are often used to categorize distinct
phases of the search process
 Exploration is
 the generation of new individuals in as-yet untested regions of the search space
 Exploitation means
 the concentration of the search in the vicinity of known good solutions
 Evolutionary search processes are often referred to in terms of a trade-off between
exploration and exploitation
 Too much of the exploitation can lead to inefficient search
 Too much of the exploration can lead to a propensity to focus the search too quickly.
 Premature convergence is the well-known effect of losing population diversity too quickly, and getting
trapped in a local optimum
Multimodal problems
 These are problems in which there are a number of points that are
better than all their neighbouring solutions
 We call each of these points a local optimum
 We denote the highest of these as the global optimum
 A problem in which there is only one local optimum is known as
unimodal
 Multimodal problems are special kind of problems where a unique
global solution does not exist
 Several global optimums or one global optimum with several local
optimums (or peaks) can be found around the search space.
Multimodal problems
 Multimodality is
 a typical aspect of the type of problems for which EAs are often
employed
 either in attempt to locate the global optimum
 or to identify a number of high–fitness solutions corresponding to
various local optimal
 Multimodal optimization
 Deals with optimization tasks that involve finding all or most of
the multiple (at least locally optimal) solutions of a problem, as
opposed to a single best solution (Wikipedia)
Characterizing Selection and Population Management Approaches
for Preserving Diversity (Mechanism to aid multimodal problems)

 These can be broadly separated into two camps:

 explicit approaches
 inwhich specific changes are made to operators in order to
preserve diversity
 and implicit approaches
 inwhich a framework is used that permits, but does not
guarantee, the preservation of diverse solutions
Diversity and Space
 Just as biological evolution takes place on a geographic surface, but
can also be considered to occur on an adaptive landscape, so we
can define a number of spaces within which the evolutionary
algorithms operate:
 Genotype Space
 Phenotype Space
 Algorithmic Space
 Diversity
 Fitness Sharing
 Crowding
 Speciation
Explicit approaches to diversity
maintenance
 Explicit approaches to diversity maintenance based on
measures of either genotype or phenotypic space
include
 Fitness Sharing
 Crowding
 Speciation

 all of which work by affecting the probability

distributions used by selection
 Fitness sharing
 It is a diversity mechanism
 The search space may includes multiple, equally optimal solutions
 To prevent the population from converging to a single optimum, fitness sharing is applied
 Fitness sharing distributes the population among multiple solutions so that only a few individuals are
maintained in the vicinity of each solution in the search space
 Based on the idea that
 individuals in the population that are close to each other have to share their fitnesses in a similar way to how species in
nature occupying the same ecological environment have to share resources
 Hence by derating the fitness of close individuals one hopes to encourage the population to spread out
more
 The main idea is to
 try and introduce niches in the population to prevent the algorithm from converging to a single solution such that different
niches explore different peaks of the fitness landscape
 Thus, in this context, niches are often understood as narrow, connected areas of the search space

Ref: https://www.sciencedirect.com/science/article/pii/S0304397518304882
Fitness Sharing
 Fitness sharing distributes the population among multiple solutions so that only a few
individuals are maintained in the vicinity of each solution in the search space
 The fitness of each individual in the current population is reduced based on the number of
similar individuals in the population
 A modified fitness value is calculated for each population individual by
 dividing the fitness, 𝐹, by the niche count of the individual, 𝑚𝑖 , which is defined by
 𝑚𝑖 = σ𝑗∈𝑝𝑜𝑝 𝑠ℎ(𝑑𝑖𝑗 )
 Where 𝑑𝑖𝑗 is the Euclidean distance between individuals 𝑖 𝑎𝑛𝑑 𝑗 and
 𝑠ℎ(𝑑𝑖𝑗 ) is the sharing function which is given by
𝑑𝑖𝑗 𝛼
 𝑠ℎ 𝑑𝑖𝑗 = ቐ1 − 𝜎𝑠
𝑖𝑓 𝑑𝑖𝑗 < 𝜎𝑠
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
 Here 𝛼 determines the shape of the sharing function. Example if 𝛼=1 then the sharing function is linear
 𝜎𝑠 is the share radius which decides both
 How many niches can be maintained
 Granularity with which different niches can be discriminated
Fitness Sharing
 Hence modified fitness value is given by
𝑭 𝒊
 𝑭′ 𝒊 =
𝒎𝒊
 Where 𝑚𝑖 = σ𝑗∈𝑝𝑜𝑝 𝑠ℎ(𝑑𝑖𝑗 )
 Where 𝑑𝑖𝑗 is the Euclidean distance between individuals 𝑖 𝑎𝑛𝑑 𝑗 and
 𝑠ℎ(𝑑𝑖𝑗 ) is the sharing function which is given by
𝑑𝑖𝑗 𝛼
 𝑠ℎ 𝑑𝑖𝑗 = ቐ1 − 𝜎𝑠
𝑖𝑓 𝑑𝑖𝑗 < 𝜎𝑠
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
 Here 𝛼 determines the shape of the sharing function. Example if 𝛼=1 then the sharing
function is linear
 𝜎𝑠 is the share radius which decides both
 How many niches can be maintained
 Granularity with which different niches can be discriminated
Crowding
 It is way of preserving diversity by ensuring that new
individuals replaced similar members of the population
 It prevents a single genotype from dominating a
population
 By ensuring that a newborn individual will replace one
that it is genotypically similar to it, Crowding tries to
maintain a balanced population
Deterministic Crowding
 This algorithm relies on the fact that offspring are likely to be similar to their
parents as follows
 The parent population is randomly paired
 Each pair produces two offspring via recombination
 These offspring are mutated and then evaluated
 The four pairwise distances between offspring and parents are calculated
 Each offspring then competes for survival in a tournament with one parent, so that the
intercompetition distances are minimized
 In other words, denoting the parents as 𝑝, the offspring as 𝑜, and using the subscript to
indicate tournament pairing, 𝑑(𝑝1, 𝑜1) + 𝑑(𝑝2, 𝑜2) < 𝑑(𝑝1, 𝑜2) + 𝑑(𝑝2, 𝑜1)
 Net result of all this is that offspring tend to compete for survival with the most
similar parent
 So subpopulations are preserved in niches but their size does not depend on fitness;
rather it is equally distributed amongst the peaks available
 Fitness sharing and crowd sharing

𝐹𝑖𝑡𝑛𝑒𝑠𝑠 𝑠ℎ𝑎𝑟𝑖𝑛𝑔 𝐶𝑟𝑜𝑤𝑑 𝑠ℎ𝑎𝑟𝑖𝑛𝑔

Ref: A.E. Eiben, J.E. Smith, Introduction to Evolutionary Computing, 2nd edition, Natural Computing Series, Springer, 2015