Week 8

Question 1

Why MOEAs?

Answer 1

Many real world problems exist with multiple objectives
Engineering problems are often multi-objective
Useful in industry to see trade-off between costs and benefits of solutions.

Question 2

How do we get EAs to perform multi-objective computation?

Answer 2

Generational GA
Revised objective function - multiple objectives
Revised selection criteria - domination criterion
Revised visualisation - pareto-optimal curves

Question 3

What is domination criterion?

Answer 3

A solution a is said to dominate another solution b if it is at least as good as b in every dimension and better than b in at least one dimension.

Question 4

What is the pareto-front?

Answer 4

Given a domination criterion, we know the best solutions will lie along a curve consisting of non-dominated points. Known as the pareto-optimal front or the pareto front.

Question 5

Desirable characteristics of the pareto front

Answer 5

Evenly spaced solutions
Covering the largest possible area of the front.

Question 6

What is discrimination in multiple objectives

Answer 6

Instead of having single fitness function, we have multiple fitnesses
Best solutions lie on pareto front
To discriminate between the rest of the solutions we use ranking methods

Question 7

Ranking method

Answer 7

Determine the pareto-front, assign it to rank 1, remove it from population, assign the next pareto front to rank 2, and so on.

Question 8

Alternative ranking method

Answer 8

Rank based on the number of solutions that dominate the solution - e.g. rank 0 = non dominated
rank 1 = one solution dominates, etc

Question 9

How do we tell the difference between two solutions which are non-dominated?

Answer 9

A new selection operator

Question 10

Problem with Pareto Domination tournament selection

Answer 10

Cannot pick winners between solutions in the same rank
Selection pressure is not sufficient, so a randomly selected comparison set is introduced.
‘Select 2 random individuals a&b and a separate comparison set c from the population. If a or b is non-dominated with respect to c, select.
If a and b have the same domination, tiebreak.

Question 11

How do we tiebreak for individuals with the same domination?

Answer 11

Choose the one from the less dense area. As it represents a wider range.

Question 12

What is niching?

Answer 12

When tiebreaking, use a niche radius around the point, the number of points within this radius is the niche count. The solution with the smallest niche count is chosen.
New parameter- niche radius.

Question 13

What is crowding distance?

Answer 13

Sum the manhattan distance of the point from the 2 nearest points.

Question 14

How does NSGAII work?

Answer 14

Initial population N
Select, crossover, mutate -> 2N.
Fast ND sort into ranks
Sort by crowding distance and discard everything beyond N.

Question 15

Elitist vs Non-elitist MOGAs

Answer 15

Elitist has superseded non-elitist.
Elitist requires no extra parameters e.g. niche sizes.
Executes and converges faster
Can prematurely converge on certain problems.

Question 16

Can we use single objective GAs for multiple objectives?

Answer 16

Yes, by weighting each objective, we can optimise for different points on the pareto curve.
You have to do a separate GA run for each point.
Final shape of the curve not defined for MOGAs.

Question 17

Computing machinery vs human ability

Answer 17

Computers outperform us in certain domains
Complex calculations,
Memory task

We can do:
Perception
Reasoning
Learning.

Question 18

Architecture of computers vs brain

Answer 18

Computers, very fast serial machines
Brain is slower but very parallel

Question 19

Key aspects of a computer

Answer 19

Can perform tasks sequentially incredibly quickly
Doesn’t make mistakes (most of the time, pentium floating point errors)
Memory and processing are separated
Large capacity of storing information indefinitely

Question 20

Key aspects of humans

Answer 20

Distributed architecture
Ability to learn
Ability to generalise
Structures responsible for memory/processing are not necessarily separated
Will forget unimportant information over time.

Question 21

Brain structure

Answer 21

100 billion neurons
Each connected to 10k others
Neurons communicate through synapses - effectively a configurable chemical junction between neurons

Question 22

What are neurons?

Answer 22

Systems for signal processing and memory
Connectionism was proposed based on research in neuroscience about how the information is processed, stored and communicated among neurons.
A brain is composed of billions of cells (neurons) which interact with each other

Question 23

What are neurons composed of?

Answer 23

Dendritic tree - receive signals
Cell body - process signals
Axon - transmit signals

Question 24

How do neurons function

Answer 24

Neuron receives electrical activity from other neurons along its dendrites
The axon produces a pulse based on the strength of the incoming pulse
This is then passed to other neurons connected to this one.

Question 25

How does the artificial neuron model real ones?

Answer 25

Activation function with a threshold which outputs 0 or 1 under certain conditions
OR functionality: threshold is >=1
AND functionality: threshold is >=2

Question 26

Problems with the original perceptron

Answer 26

Could not correctly solve XOR function.

Question 27

How was the issue of computing XOR solved

Answer 27

Introducing hidden layer to make MLP