Introduction to Genetic Algorithms 2014

Survival of the Fittest
An Introduction to Genetic Algorithms and
Evolutionary Computation
Aleksander M. Stensby
Monokkel A/S
03.10. 2014

aleksander@monokkel.io
CEO
/
Co-‐Founder
@astensby

Outline
Part 1: Theory lecture – Introduction to Genetic Algorithms
10:15 – 13:00
Part 2: Workshop – Traveling Salesperson Problem (TSP)
13:00 – 15:00

Lecture
outline
● An NP-Complete Problem – The TSP
● Darwin's Theory of Evolution
● Genetic Algorithms (GA)
● Applications of GA
● Genetic Operators
● Generic GA
● Why does GAs work?

Traveling
Salesperson
● Given a list of cities to visit.
● Goal: find the shortest tour that visits each city exactly once,
returning in the end to the starting point.
● Complexity: O(n!)
● NP-Hard

Darwin’s Theory of Evolution

Darwin's
Theory
of
Evolu;on
● All life is related and has descended from a common ancestor.
● Natural selection
– “Survival of the fittest”
● Organisms can produce more offspring than their surroundings
can support -> natural struggle to survive.
● Organisms whose variations best adapt them to their
environments survive, the others die.
● Variations are heritable -> can be passed on to the next
generation -> i.e., evolution

Inheritance
Muta=on
Selec=on
Crossover
Genetic Algorithms

What
is
Gene;c
Algorithms?
● John Holland (70's)
● Nature’s mechanism for evolution could be modeled in
computers to find successful solutions for difficult problems.
● By taking a population of possible answers and evaluating them
against the best possible solution, the fittest individuals of the
population are determined.
● After evaluation, combining and mutating, the members of the
current generation generate a new population.
● This new generation is then evaluated and the process is
repeated, until an optimal solution is found.

TSP
cont.
What was our TSP problem?
- population of possible answers:
Possible tours: “1-3-5-6-7-4-2-1”
- evaluate – best possible solution:
Shortest tour!
- generate a new population by combining and mutating
- evaluate new population, “rinse, repeat”

TSP
search
space
500 cities on a circle – simple?
Possible solutions?
500!

Applica;ons
● What problems can we solve with a GA?
– Optimization & Design
– TSP, function optimizations, time tables..
– Approximate NP-Hard problems
– Simulation
– Modeling, system identification
– Evolutionary machine learning
“Mom and dad jet engine can get together and have baby jet
engines. You find the ones that work better, mate them, and
just keep going.” - Goldberg

Example:
Shape
op;miza;on
● NASA: Satellite truss or boom design
– the design of satellite trusses with enhanced vibration
isolation characteristics
– produced using Genetic Algorithm methods and a highly
customized vibrational energy flow code
● Evolutionary Design: 20.000% better!!!

Example:
Antenna
design
(NASA)
● Encode antenna structure into a genome
● Use GA to evolve an antenna
● Evaluation: Convert the genotype into an antenna structure
● Simulate using antenna simulation software

GA
terminology:
● Population
● Individuals – Chromosomes – Representation?
● Generations – Evolution
● Fitness – How “fit” is the individual?
● Development – Selection – Reproduction

Evolu;on
–
from
one
genera;on
to
the
next
● Duplication? -> No improvement
● Randomly produced? -> Past advances are not preserved
● Fitness is not preserved by duplication
● Observed variety is not due to random variation
● So, how do we retain past successes?
● How do we use them to increase the probability of fit (and
novel) variants?
● Fitness proportional reproduction & genetic operators

What
should
our
GA
do?
● Recombine 'surface' similarities among the fittest individuals...
● Combine 'good ideas' from different good individuals...
● ... because certain substructures in good individuals cause their
high fitness, and recombining such 'good ideas' may lead to
better individuals...

Gene;c
operators
● Fitness-proportional reproduction
● Genetic recombination -> Crossover
● Mutation -> “copy errors” -> additions / deletions of base pairs

Gene;c
operators:
-‐
Crossover
● Crossover
– Sexual reproduction (pass on 50% of your genes)
● Benefits?
– Stability – occurs between very similar DNA segments
– Leads to clearly defined species!
– Stability - lengths of DNA molecules are preserved
– Variability – combining “good” ideas

Gene;c
operators:
-‐
Crossover
● Crossover
– Single Point
– Two point
– Uniform

Gene;c
operators:
-‐
Muta;on
● Mutation
– Insert / Delete / Substitute
● Mass mutation -> harmful! (e.g. genetic disorder)
● Small changes -> beneficial! -> Variations!
● Help to better adapt to changes in their environment!

Gene;c
operators:
-‐
Muta;on
● Mutation
– Inversion
0 0 1 1 1 0 1 0 0 1 => 0 0 1 0 1 0 1 0 0 1
– Substitution
1 2 3 4 8 7 6 9 5 => 1 2 9 4 8 7 6 3 5
– Update
8.3 1.2 4.3 2.2 2.7 7.1 => 8.3 1.2 4.3 2.3 2.7 7.1
– Insertion
1 2 3 4 8 7 6 9 5 (select random)
1 2 3 4 9 8 7 6 5 (insert at random location)

Gene;c
operators:
-‐
Reproduc;on
● Fitness-proportional reproduction / Selection strategies
● Again; survival of the fittest
● Population fitness F = Σk=1
popSize fk
● Roulette Wheel selection
● Rank selection
● Tournament selection
● Elitism
– First copies the best chromosome (or a few best chromosomes) to new
population. The rest is done in classical way.
– Elitism can very rapidly increase performance of GA, because it prevents
losing the best found solution.

RouleOe
Wheel
selec;on
- Rank individuals in increasing order of fitness,
from 1 to popsize (n)
- Probability of selecting individual vi = fi / F
for (int k = 0; k < population.size(); k++) {
sum += (population.get(k).getFitness() / populationFitness);
if (sum >= random)
return population.get(k);
}

Rank
selec;on
- Rank individuals in increasing order of fitness, from 1
to popsize (n)
- Better when fitness differs a lot
=> No super individuals
for (int k = 0; k < population.size(); k++) {
double pk = Math.pow(selectionPressure, k + 1);
sum += pk;
if (sum >= random)
return population.get(k);
}

The
components
of
a
GA
● Representation / Encoding of a Chromosome
– Binary, Permutation, Value...
● Initialization
● Evaluation / Fitness function
● Genetic operators / Selection
● Parameters
– Population size
– Xover probability
– Mutation probability
– ...

Generic
GA
–
Pseudo
code
● 1. Choose initial population - random
● 2. Evaluate the fitness of each individual in the population
● 3. Repeat until termination
● Select best-ranking individuals to reproduce (parents)
● Breed new generation through crossover and mutation (genetic operations) and give
birth to offspring
● Evaluate the individual fitness of the offspring
● Select individuals for next generation

Some
recommenda;ons
● “Generally good parameters”
– High crossover probability! (≈ 0.6)
– Low mutation probability! (≈ 0.1 to 0.001)
– Population size? - usually bigger is better!
● Chromosome / String size -> determines search space!
– e.g. 30 bits? -> search space = 230 = 1.07 billion points

Problems
with
GAs
● No convergence guarantee
● Premature convergence
● Disadvantages:
– May be difficult to choose encoding
– May be difficult to define the fitness function
– May be slow (not really a problem with todays computers)

Why
does
GAs
work?
● Directed and stochastic search!
– Population of potential solutions (randomly spread out)
– “Re-use” relatively good (surviving) solutions
– Exchange information among these relatively good solutions
– Search in multiple directions – in parallel!
● Exploration & Exploitation
● Start with an “open mind” - decisions based on randomness
– All possible search pathways are theoretically open to a GA
– “Uncover solutions of startling and unexpected creativity
that might never have occurred to human designers”
● Once you have your GA; simple to solve new problems!!

Other
evolu;onary
methods
● Genetic Programming
● Swarms / Ants
● ALife
● More advanced GAs (hierarchical GAs, Evolution strategies, etc.)

Gene;c
Programming
(GP)
1) Generate an initial population of random compositions of the functions
and terminals of the problem (computer programs).
2) Execute each program in the population and assign it a fitness value
according to how well it solves the problem.
3) Create a new population of computer programs.
i) Copy the best existing programs
ii) Create new computer programs by mutation.
iii) Create new computer programs by crossover(sexual
reproduction).
4) The best computer program that appeared in any generation, the best-so-
far solution, is designated as the result of genetic programming
Example: http://genetic.moonlander.googlepages.com/

Demonstra;ons
● GA & Music, Art
● http://jgap.sourceforge.net/
● http://kandid.sourceforge.net/
● Spore, anyone? - Evolving creatures! Karl Sims
● Lee Graham
● http://www.stellaralchemy.com/lee/virtual_creatures.html
● http://www.youtube.com/watch?v=l-qOBi2tAnI
● http://www.youtube.com/watch?v=F-GnKr4rw4M
● http://www.youtube.com/watch?v=OxK5OFPOMZU
● http://www.youtube.com/watch?v=25fFoFxYg7o
● http://www.youtube.com/watch?v=kSXeqPbAP5I
● http://www.youtube.com/watch?v=O82tVjDBc7w
● http://www.youtube.com/watch?v=U5GqpH6EZvo

Workshop
- Representation: Permutations
- Crossover:
● PMX (Partially Mapped Xover)
● Order Xover
● Position Based Xover

Workshop
- PMX (Partially Mapped Xover)

Workshop
- Position-Based Xover

Workshop
• Mutation:
• Inversion
1 2 3 4 5 6 7 8 9 ==> 1 2 6 5 4 3 7 8 9
• Insertion
1 2 3 4 5 6 7 8 9 (select random city) ==>
1 2 6 3 4 5 7 8 9 (insert in random spot)
• Reciprocal Exchange
1 2 3 4 5 6 7 8 9 ==> 1 2 6 4 5 3 7 8 9

Bibliography
● “Adaption in Natural and Artificial Systems”, Holland, 1975
● Evolutionary Computation, Lecture notes, F. Oppacher,
Carleton U.
● http://www2.econ.iastate.edu/tesfatsi/holland.gaintro.htm
● http://ti.arc.nasa.gov/projects/esg/research/antenna.htm
● http://www.talkorigins.org/faqs/genalg/genalg.html
● http://www.obitko.com/tutorials/genetic-algorithms/

Introduction to Genetic Algorithms 2014

More Related Content

What's hot

Similar to Introduction to Genetic Algorithms 2014

Recently uploaded

In this document

Introduction to Genetic Algorithms 2014