OUTLINE
• WHAT ARE EVOLUTIONARY ALGORITHMS
• HOW CLASSICAL GENETIC ALGORITHMS WORK
• FROM GA TO DIFFERENTIAL EVOLUTION
• PARTICLE SWARM OPTIMIZATION
Darwin’s core statement in his own language (1859)
“As many more individuals are produced than can possibly survive,
there must in every case be a struggle for existence, either one
individual with another of the same species, or with the individuals of
distinct species, or with the physical conditions of life… . Can it, then, be
thought improbable, seeing that variations useful to man have
undoubtedly occurred, that other variations, useful in some way to each
being in the great and complex battle of life, should sometimes occur in
the course of thousands of generations? If such do occur, can we doubt
(remembering that many more individuals are born than can possibly
survive) that individuals having any advantage, however slight, over
others, would have the best chance of surviving and of procreating their
kind? On the other hand, we may feel sure that any variation in the least
degree injurious would be rigidly destroyed. This preservation of
favorable variations and the rejection of injurious variations, I call
Natural Selection.”
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
WHAT ARE EVOLUTIONARY ALGORITHMS?
Application of the principle of Natural Selection –
The Survival of the Fittest – to the problem of
Optimization.
HOW?
Darwin was unaware of Genetics
• Darwin was unaware of the existence of genes,
chromosomes or DNA
• However, he with his team was aware that parents
somehow pass on their physical characteristics to children
• Gregory Mendel, Austrian monk (1866), postulated the
theory that small elements in the cells of parents codify
these physical characteristics and pass them on to
children as pairs – one from each parent
• His theory became popular only after 1900 when
scientists actively starting researching into and
developing this field
• Now Gregor Mendel is known as the father of Genetics.
From Natural Selection of Species – through Natural
Selection of Chromosomes – to Natural Selection of
Optimal Solutions
• It was not until the 1960’s that people started thinking –
• Can not the genetic techniques that enable the natural
evolution of biological species …..
• ….. be applied to the natural evolution of optimal solutions
….
• if these artificial optimal solutions can somehow be
expressed as chromosomes and genes?
• In the late 1960’s and early 70’s , Ingo Rechenberg in Germany
and John Holland in USA independently came up with methods for
the artificial evolution of optimization problems
• Rechenberg came up with Evolutionary Strategies and Holland
with Genetic Algorithms
• Other techniques were also invented that broadly come under
the umbrella of Evolutionary Algorithms.
Biologically Inspired Algorithms
• There is an important class of Algorithms in Artificial
Intelligence that are known as Biologically Inspired
Algorithms
• They Include:
• Evolutionary Algorithms
• Swarm Algorithms (also known as Socially Inspired
Algorithms)
• Fuzzy Logic
• Artificial Neural Networks
• They comprise an important subset of AI called
Computational Intelligence and also Soft Computing.
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
In the N- dimensional (assuming N- parameters) solution space there
exists one point that is the optimum – the global optimum
At this point the objective function which we seek to optimize reaches a
global maximum or minimum value
Every single point in this space is associated with a “fitness”, i.e. where it
stands relative to this global optimum; there is a one-to-one mapping
between the location of each point and its fitness
A number of candidate solutions – the “population” – are created that
concurrently traverse the total solution space across “generations” –
striving to “evolve” to higher and higher levels of fitness
When a candidate moves from one point to another (evolves), what
essentially changes is its state, i.e. the values of its component variables –
very often coded as bits called the genes that compose the chromosome
When moving from one generation to the next, the fitter solutions are
reproduced more, while the weaker ones are gradually eliminated
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
CLASSICAL GENETIC ALGORITHMS
SELECTION: More the fitness, more the probability of being
reproduced into the next generation
CROSSOVER: A pair of candidate solutions flip their genes within
a crossover zone
A A A A A A A A A A A A A A A
CROSSOVER POINT
B B B B B B B B B B B B B B B
AFTER CROSSOVER:
A A A A A A A A B B B B B B B
CROSSOVER POINT
B B B B B B B B A A A A A A A
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
CLASSICAL GENETIC ALGORITHMS
• Note that Stochastics (Randomness) plays a crucial role in ALL
Evolutionary Algorithms.
• Thus which pairs are crossed over are selected randomly
• Importantly, there is a probability of crossover (say 0.7), and this
is effected by generating a random number for each (randomly
selected) pair and checking if that number is within the CP
• the point of crossover in the chromosome is also selected
randomly.
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
CLASSICAL GENETIC ALGORITHMS
MUTATION: An “Arbitrary” bit in an “Arbitrarily” selected
Chromosome is flipped – role of stochastics and
probability of mutation
SELECTION: More the fitness, more the probability of being
reproduced into the next generation. Roulette wheel
selection and the role of stochastics.
ELITISM: In a new generation, after crossover & mutation, if the
fitness of the best solution has deteriorated from the
best fitness of previous generation, then the worst
solution of current generation is replaced by that best
solution (of previous gen.)
Roulette Wheel Selection
CROSSOVER – MATHEMATICAL EXPRESSION
Formally, if the dimensionality of the solution space is denoted as
D and the number of candidate solutions is N, then the elements
of the ith solution vector Xi,G at generation G may be denoted as
X i ,G ( x1,i ,G , x2,i ,G , x3,i ,G , ........, xD ,i ,G ) i {1, 2, ..., N}
new
x
Upon crossover between two candidates i and k, an element
j ,i ,G
, i.e. element j of the new candidate i is given as
new
x j ,k ,G if CrP j D where CrP denotes
random x
j ,i ,G
x j ,i ,G otherwise
Crossover point; 1 CrP
D x new
j , k ,G
x
j ,i ,G if CrP j D
while new
an
x
j , k ,G
element , i.e. element j of the new candidate k is
given as x
j , k ,G otherwise
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
Evolutionary Algorithms have two advantages over
gradient-based methods:
- They are unlikely to get stuck in local optima
- They can be applied to problems that are not analytic, and even
those that are not mathematically tractable – as long as you can
measure the fitness of a candidate
They differ from one another essentially in the mechanism
by which a candidate solution evolves from one state (i.e.
one point in the solution space) to the next
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
CLASSICAL GENETIC ALGORITHMS
A SHORTCOMING: The GA brings the solution within a close
neighborhood of the global optimum, and then,
instead of moving straight towards that point , it
tends to “beat around the bush”
ACCELERATION SCHEMES: Various techniques have been developed to
accelerate the solution in this neighborhood zone,
and one class of these techniques plays with the
mutation
ALTERNATE APPROACHES: Keeping in mind the pros and cons of
classical GA, alternate approaches founded
on the evolutionary principle have been
developed. One such approach is
Differential Evolution
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
DIFFERENTIAL EVOLUTION RAINER STORN, 1995, Univ. of Calif, Berkeley
STEP 1: CREATE A NEW SOLUTION CALLED MUTANT VECTOR AS
xi , g xi , g K *( xr1, g xi , g ) F *( xr 2, g xr 3, g )
where i, r1, r2, r3 are all different, randomly selected
members of the chromosome pool (i.e. population) in
generation “g”,
K and F are relaxation parameters
- in effect a mutant vector is being created for member “i” by
superposing on the old solution suitably scaled differences between
randomly selected alternate members of the population
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
DIFFERENTIAL EVOLUTION: CROSSOVER
STEP 2: Use the Mutant Vector as a crossover partner for the old
solution to generate the new solution; i.e. some randomly
selected variables composing the old solution are replaced by
corresponding variables from the mutant vector to create the
new solution – called the TRIAL VECTOR
Crossover is performed between the ‘mutant’ vector Vi,G and the
existing (parent) vector Xi,G to generate a ‘trial’ vector Zi,G
according to
v j ,i ,G if rand j (0,1) Cr
z j ,i ,G
x j ,i ,G otherwise
where zj,i,G is the element j of the trial vector Zi,G, randj(0, 1) denotes a
random number between 0 & 1 applied to the element j, Cr is the
crossover threshold (probability).
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
DIFFERENTIAL EVOLUTION: ELITISM
STEP 3: Check if the new solution member “i” has higher fitness than
the previous one; if yes, only then replace:
this is Elitism (vertical).
Question related to Step 3: Where is the difference with GA and what
are the possible implications?
IT IS SEEN THAT DE’s GENERALLY LEAD TO FASTER CONVERGENCE
THAN CLASSICAL GA’s – SEE ILLUSTRATION IN NEXT SLIDE.
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
DIFFERENTIAL EVOLUTION RAINER STORN, second-half-90’S
xi , g xi , g K *( xr1, g xi , g ) F *( xr 2, g xr 3, g )
Close after Start
FROM GENETIC ALGORITHMS TO DIFFERENTIAL EVOLUTION
DIFFERENTIAL EVOLUTION RAINER STORN, second-half-90’S
xi , g xi , g K *( xr1, g xi , g ) F *( xr 2, g xr 3, g )
Close to convergence
Implementation of DE Solution
Initial
Population
START
For a Generation g Evaluate Cost
For a candidate in
Population p Selection using Elitism
g = g+1
Generate mutant
vector No Traversed all
candidates ?
Generate Trial Vector Yes
P = p+1
No Solution
Evaluate Constraints converged ?
Yes
Constraints Identify best member
No satisfied ?
in pool
Yes
END
Algorithm for Design and Implementation of a DE Solution
• Step 0: Get the Inputs: the number of design variables for
your system, the range of each variable, the objective
function, the constraints
• Step 0A: Normalize the input variables only if the nature of
the objective function does not get altered by the
normalization
• Step 1: Decide on the meta-parameters: population size,
crossover probability, value of two mutation parameters K
and F, max. number of generations
• Step 2: Initialize all candidates of the population. Simplest
approach: take all random. Better approach: ensure that they
are spread over the entire solution space
• Step 2A: Confirm that all constraints are satisfied by each
Algorithm for Design and Implementation of a DE Solution
• Step 3: Go over each generation
• Step 4: Go over each member of population (i.e. candidate)
• Step 5: Randomly select the candidates to mutate with this
candidate
• Step 6: Perform mutation to generate the Mutant Vector
• Step 7: Perform crossover to generate the Trial Vector
• Step 8: Check the Trial Vector for satisfaction of all
constraints, if it fails, repeat steps 5-7 till it passes
constraint checks
• Step 9: Compare fitness of Trial Vector with Parent Vector
and select the fitter one to get into the next generation
• Step 10: End population loop
• Step 11: Check if solution converged (your convergence
Particle Swarm Optimization
Particle Swarm Optimization
• Particle Swarm Optimization (PSO) is the foremost among
the Swarm Optimization Algorithms, which can be called
Cousins of Evolutionary Optimization Algorithms
• The similarity is that both work upon a population of
candidate solutions that improve in parallel across a large
sequence of iterative steps
• Dissimilarity: while Evolutionary Algorithms create new
solutions from old ones by combining them through
crossover- and mutation- like (Genetic Evolution)
operations, in Swarm Algorithms the candidates maintain
their individual identities and travel over the parameter
space by exchanging information with each other (like
Swarms)(no “Selection” step)
• Strictly speaking, Swarm Algorithms can be more
Particle Swarm Optimization
Particle Swarm Optimization
• Let us take N candidates of a population in a multi-
dimensional (M) solution space, we will call these
candidates as “particles”
• At any iteration G – strictly we no longer call them
generations but instead iterations, however, for similarity
with earlier formalisms we still call it G – we have these N
particles at various positions pi and moving with
corresponding velocities vi, for i {1, 2, …, N}
• Importantly, each particle retains memory of its own best
ever position pi,best – where best-ever implies that position
which had the best fitness when mapped into objective
space – and has information also of the gbest, i.e. the global
best ever attained by any particle
Particle Swarm Optimization
• Then the velocity and position update rules of particle i, in
the classical PSO Algorithm, is
viG 1 viG c1 rand ( pi ,best piG ) c2 rand ( gbest piG )
and
piG 1 piG viG 1
• where, apart from the earlier defined notations,
• c1, c2 are the weights of local and global information
• rand is a random number in [0, 1], different for each use
• it is being assumed that velocity represents the change in
position from one iteration to the next
• c1 and c2 are empirical parameters (meta-parameters) but these
are selected by rule of thumb such that c1 + c2 = 4
• higher velocities lead to faster movements but with higher
chances of instability, hence a max. vel. vmax is set for each
dimension
Particle Swarm Optimization: Algorithm
for each particle
initialize particle # Try to ensure maximum spread across
solution space
if constraints not satisfied, re-initialize
obtain fitness
fbest, pbest fitness, position
broadcast fbest, pbest
gbest pbest of best(all fbest)
v 0
end
do # for iteration or “generation” loop
for each particle
evaluate new v # Check and implement bound
evaluate new p
evaluate fitness
if fitness < fbest # for minimization problem, else
reverse
fbest, pbest fitness, p
broadcast fbest, pbest
gbest pbest of best(all fbest)
THANK YOU
