1. Evolutionary Computation
Genetic Algorithms

2. Outline Evolution Genetic Algorithms
Selection, inheritance, variation (mutation)
Natural Selection
Artificial Selection
Some examples – successes and problems.
Genetic Algorithms
Selection
Mutation
Crossover
Fitness evaluation

3. Evolution – Darwin and Mendel

4. Survival of the fittest
“white” moths are difficult to see against some trees.
Which moth survives?
Then there was the industrial revolution – and tree turned black with “soot”.
Which moth survives now?

5. Generate and Test Approach
A “solution” is generated.
It is tested on a problem instance
Each solution is assigned a score (real value).
This process is repeated until
time expires or a solution is found.
Includes much of machine learning.
We try to improve the quality of solutions generated by using feedback in this loop.
Alternative but equivalent view is we are sampling a space.
test
Generate
test
Test
9/22/2018
Automatically Designing Selection Heuristics

6. Generate and Test Examples
GECCO 1st workshop on Evolving Generic Algorithms.
Generate and Test Examples
Manufacture e.g. cars
Evolution “survival of the fittest”
“The best way to have a good idea is to have lots of ideas” (Linus Pauling).
Computer code is also generate and tested.
9/22/2018
Automatically Designing Selection Heuristics

7. Automatically Designing Selection Heuristics
Fit For Purpose
Evolution “designs/generates” organisms for a particular environment.
Similarly we should design metaheuristics for particular problem class.
“we propose a new crossover operator…”
“what is it for…”
9/22/2018
Automatically Designing Selection Heuristics

8. Evolution Evolution has 3 components.
Selection “survival of the fittest”.
Inheritance (i.e. your genes are inherited from your parents, and you look similar to your parents).
Variation or mutation, random changes to your DNA

9. Examples of Evolution 1
E. coli reproduce copy DNA in 40 minutes but can give birth every 20 minutes – how ???.
Hearing in 3D – human ears – how ???
The eye – design around 40 different ways.
Mankind – does not mean “man”-kind but “manipulator” kind

10. Humans
What makes humans unique compared to other animals? What make us so much more intelligent?
Hands – compare with the apes
Brains – large & expensive
We walk “upright” – is this important
Our diet – more complex than other animals
Language (spoken) – instant communication
Language (written) – communication over a period of time.
Mirror Test and self awareness.
What other animals can pass the mirror test???

11. Mirror Test
The mirror test is a measure of self-awareness, as animals either possess or lack the ability to recognize themselves in a mirror.
Animals that have passed the mirror test include: all of the great apes (bonobos, chimpanzees, orangutans, humans, and gorillas), rhesus macaques, bottlenose dolphins, orcas, elephants, and European Magpies.
Robots??? Young babies.
Mirror neurons.

12. Examples of Evolution 2 Giraffe laryngeal nerve No wheels in nature?
Bulldogs – big head – one problem ???
Whales
taste like beef
have bones from back legs
Wave tail vertically.

13. Artificial Selection – Selective Breeding
Instead of nature deciding which individuals in a species lives or dies, we (humans) can decide.
Roses selected for beauty (appearance not smell)
Cows selected for milk production
Horses (race horses, carthorses)
Sheep selected for wool production
Dogs selected for ???.

14. Evolutionary Computation
How can we use the ideas of evolution to solve problems on a computer?
Genetic Algorithms evolve a population of bit strings (bits can represent on/off choices or numbers)
Genetic Programming evolves a population of programs used for e.g. controlling an agent in a game, a robot in a maze….

15. Example Problem
The knapsack problem or rucksack problem is a problem in combinatorial optimization
Given a set of items, each with a weight and a value, determine the number of each item to include in a collection so that the total weight is less than or equal to a given limit and the total value is as large as possible.
It derives its name from the problem faced by someone who is constrained by a fixed-size knapsack and must fill it with the most useful items.
maximize
Subject to the constraint

16. Bit String Representation
Items 1, 2, …i,…,n
Item i has weight wi and value vi
What is the time complexity to try all combinations?
We can use a binary representation e.g.
(01…) mean do not include item1, do include item 2
0=do not include, 1= include.
A genetic algorithm will sample a subset of solutions and hopefully find a good enough solution.
There is no guarantee that the optimal solution will be found.

17. Fitness function 10110 has a value 4+0+1+10+0=15$2 3 4 5
VALUE $ 10
WEIGHT KG 12
10110 has a value =15$
And has weight = 17kg
But the bag has capacity 15kg – so this solution is not feasible/is invalid (give value -1)
Fitness(10110) = -1
10100 has a value =5$
This is okay (<15kg) so we can give it a fitness of 5$
Fitness(10100)=5$

18. How to solve it?
Enumeration – list all possibilities (how many are there?), and try each one.
What is the problem with this approach?
A second way is to randomly generate bit-strings and find the value of each one.
What are the problems with this approach.
Neither of these methods are very good.

19. Basic Outline/Algorithm
Repeat n times
Select pair of bit-strings from population
Perform crossover to produce 2 children.
Mutate them
Calculate their fitness (also called utility) and return them to population.

20. Selection Two examples Rank selection Fitness Proportional selection1 2 3 4
Bit string
0011
1010
1110
0001
fitness 5 10 20
Two examples
Rank selection
Individuals are selected in proportion to their rank in the sorted population
Fitness Proportional selection
Individuals are selected in proportion to their fitness in the sorted population.
If we monotonically scale the fitness, there is not difference to rank, but there is a change to fitness.

21. Crossover Two crossover One point crossover 1111 and 0000 gives ???
Two point crossover

22. Mutation Two examples One point mutation Uniform mutation
only flip one bit e.g > 1010
What else is possible ???
Uniform mutation
Flip all bits with e.g probability e.g > 1110.

23. Fitness
We could give a bit string the fitness of the direct value it has
E.g. value = 9 therefore fitness = 9
Or we could scale it (most books do not mention this)
Ax+B where x is the fitness
Or any monotonic function
Or possibly a non-monotonic function.

24. No Right or Wrong way There is not right way to do this.
For selection/ crossover/ mutation / fitness I have told you a choice of two for each – that is 8 possible ways.
You can try to find your own ways.
This is a very large area of research.
Goldberg: 8th most referenced book in computer science last century.
Russell and Norvig is 2nd most referenced computer science book this century.
Youku.com

25. Evolutionary Video A video about walking robots (animation)
(time = 50:00)