Lecture 3 -Evolutionary Computation

Question 1

Which mechanisms are at place between parents and offspring generation (2)?

Answer 1

1. Recombination

2. (random) Mutation

Question 2

First step of designing EA?
+ what does it need to be?
+ What forms can it take?

Answer 2

Designing a representation (phenotype) for an individual solution
+ that must be relevant to the problem we’re solving
+ it can be discrete value (binary, integer)

Question 3

Example discrete representation

+ example mutation
-> how often does it happen?

Answer 3

Binary alphabet
10100011
-> single 0’s are genes, the whole 8 bit block is a chromosome

Mutation
10101011
-> based on certain probability for each gene

Question 4

Example order based representation
+ What are individuals represented as?

+ What is it used for?

+ Example for mutation

Answer 4

Traveling salesman (city gets assigned number 1-N, solution could be 5,4,2,1,3 -> fasted way)

+Individuals are represented as permutations

+Used for ordering and sequencing problems

original 
1 2 3 4 5
mutation 
1 4 3 2 5
-> Swap

Question 5

What is the most costly step for a real EA application?

-> What do we have to avoid?

Answer 5

Evaluating an individual

-> don’t re-evaluated unmodified individuals

Question 6

Mutation operators

1. One or many?
2. What should it be able to reach?
3. Size (2)
4. Solutions

Answer 6

1. At least one
2. Should allow every part of the search/pattern space to be reached
3. The size of mutation is important and should be controllable
4. The mutation should produce valid solutions/chromosomes

Question 7

Recombination operators

(3) important points
2. representation?

Answer 7

1. Child should inherit from each parent
2. The recombination operator should be designed in conjunction with the representation
3. Recombination should produce valid chromosomes

Question 8

Example Recombination for discrete representation

Answer 8

Parent 1   Parent 2 
1 1 1 1 1      0 0 0 0 0 
cut at random point (n=3)
child 1       child 2
1 1 1 0 0    0 0 0 1 1

Question 9

Selection strategy

1. What do we want to ensure?
2. What will this give us?
3. What do we have to be careful of?

Answer 9

1. Ensure that better individuals have higher chance of being parents
2. this creates selection pressure
3. less good individuals need some chance of being parents too - they might include useful genes

Question 10

Example rank based selection

1. What are individuals sorted on? What is this place called?
2. How are individuals selected?
   - > example

Answer 10

1. Sorted according to fitness value. The place in the list is called a rank.
2. This rank is then used by a function that selects individuals. Individuals with higher ranks have a higher chance of being selected
   - > roulette

Question 11

Replacement strategy

1. what does it affect? How so?

Answer 11

Deciding which individuals to kill effects the selection pressure

Question 12

Elitism

Answer 12

Re-introducing the individual with the highest fitness in the population

Question 13

Preservation

Answer 13

Keep individual with highest fitness in safe place

Question 14

Recombination vs Mutations

1. Recombination
   - what does it depend on?
   - what changes effect
   - ..

Answer 14

• modifications depend on the whole population
• it’s effect decreases with convergence
• exploitation operator

Question 15

Recombination vs Mutations

2. Mutations
   - what is it good for?
   - strong…
   - …

Answer 15

• Escaping the local optima
• it has a strong causality principle
• exploration operator

Question 16

Stopping criterions (4)

Answer 16

1. The optimum is reached
2. Limit on CPU resources (there is a max amount of fitness evaluations)
3. User’s patience
4. Some generations without improvements

Question 17

Algorithm performance

1. Design perspective
2. Production perspective
3. What should we never do?

Answer 17

1. Find a very good solution at least once
2. Find a good solution at almost every run
3. Never draw conclusion from one run (use stats)

Question 18

Practical performance

1. Acceptable performance at … on …
2. Intrinsic…

Answer 18

1. Acceptable performance at acceptable cost on a wide range of problems.
2. Intrinsic parallelism (robustness, fault tolerance)

Question 19

At which problems is EA superior (3)?

Answer 19

1. Lots of data with many free parameters
2. Complex relationship between parameters
3. Many (local) optima

Question 20

Key issue

1. Exploration

Answer 20

Sample unknown regions

If too much exploration (random search) there will be no convergence

Question 21

Key issue

2. Exploitation

Answer 21

Try to improve the best-so-far individuals
Too much of it leads to local search only.
The algorithm will converge to a local optimum.