Week 8 - Multi-Objective EAs and Neural Networks

Question 1

What is domination criterion?

Answer 1

A solution a is said to ‘dominate’ another b in the population if it is at least as good as b in every dimension and better than b in at least one dimension (objective).

Question 2

What is the pareto-front?

Answer 2

Given the domination criterion, we know that the best solutions will lie along a curve consisting of non-dominated points. This is known as the ‘pareto-optimal front’ or ‘pareto-front’.

Question 3

Name two desirable characteristics of the pareto-front.

Answer 3

Evenly-spaced solutions and covering the largest possible area of the front.

Question 4

In multi-objective EAs, where does the best solution lie?

Answer 4

On the pareto-front.

Question 5

How does an optimal pareto domination tournament work?

Answer 5

Select two random individuals a & b and a separate comparison set c from the population. If a or b is non-dominated with respect to c, then select. If a and b have the same domination – tiebreak.

Question 6

What is niching?

Answer 6

Separate fitness landscape or genotype into ‘Niches’. Prefer individuals in a niche with less other individuals.

Question 7

What does NPGA stand for?

Answer 7

Niched Pareto Genetic Algorithm

Question 8

What does NSGA stand for?

Answer 8

Non-dominated Sorting Genetic Algorithm

Question 9

What is a neuron?

Answer 9

A system for signal processing and memory. A brain is composed of billions of neurons.

Question 10

What are the three parts that make up a neuron?

Answer 10

Dendritic tree (receives signals), cell body (processes signals), axon (transmit signals)

Question 11

Describe how a neuron functions.

Answer 11

A neuron receives electrical activity from other neurons along its dendrites (inputs).

The cell body makes a non-linear decision based on incoming pulses.

The axon will produce a pulse based on the strength of the incoming pulse.

It is then passed to other neurons connected to this one via a synapse.

Question 12

What is a synapse?

Answer 12

A chemical junction which can be modified and is thought to be where the learning takes place. Synapses can excite or inhibit the dendrite. The synapse can release more neurotransmitter to enhance the coupling between cells.

Question 13

What did McCulloch and Pitts create in 1943?

Answer 13

The artificial neuron, capable of processing simple logical expressions.

Question 14

What is Hebbian Learning (1949)?

Answer 14

1. When two neurons fire together the connection between neurons is strengthened.
2. The activity of firing is one of the fundamental operations necessary for learning and memory.

Question 15

What did Rosenblatt create in 1962?

Answer 15

The perceptron, which can learn by means of weighted connections.

Question 16

Name a problem with Rosenblatt’s perceptron?

Answer 16

It could not correctly solve the XOR function.

Question 17

What did the multi-layer perceptron solve?

Answer 17

The XOR problem a single perceptron could not solve.

Question 18

Name some applications of neural computing.

Answer 18

Image analysis (e.g. facial recognition)
Autonomous vehicles
Robotics
Data analysis
Classification/prediction
Pattern detection
Control systems in engineering applications (e.g. power stations and other safety critical systems)

Question 19

Name some variations on the perceptron.

Answer 19

Multi-Layer Perceptrons (including deep neural networks)
Recurrent neural networks
Self-organising maps
Kohonen networks
Boltzmann machines
Probabilistic neural networks

Question 20

What does MOGA stand for?

Answer 20

Multi Objective Generational Algorithm

Question 21

How does NPGA work?

Answer 21

Generate initial population of solutions

Select individuals:
1. Run pareto-tournament domination selection. If one solution is non-dominated and another not, go to 3, otherwise go to 2
2. Compute niched fitness to separate solutions
3. Repeat for n selected individuals

Crossover and mutate to generate new population

Question 22

What is NSGA?

Answer 22

Elitist MOGA, generates new solutions using selection, crossover and mutation. Uses a fast non-dominated sort to rank solutions. Uses crowding distance as a tie breaker.

Question 23

What is crowding distance?

Answer 23

For each objective m:
1. Sort population by value of m
2. For end solutions crowding distance = infinity
3. For other solutions (i), the crowding distance for objective m = (i+1).m - (i-1).m and add this solution to the sum for i

Repeat for all i and for all m