Summary & Questions Weria and Kai I NF3490/ 4490 Exam Format: - - PowerPoint PPT Presentation

INF3490/4490 Biologically inspired computing

Summary & Questions

Weria and Kai

I NF3490/ 4490 Exam

• Format: Written/Digital (see small example at

uio.inspera.no)

• When: November 30, at 09:00 (4 hours)
• “Closed book exam”: No materials are

permitted on the exam

• Location: See StudentWeb and

http://www.uio.no/studier/emner/matnat/ifi/INF3490 /h18/eksamen/index.html

• http://www.uio.no/studier/emner/matnat/ifi/INF4490

/h18/eksamen/index.html

• Same exam in INF4490 as in INF3490

Multiple-choice Questions on Parts of the Exam

3

INF3490/4490 — Biologically Inspired Computing November 30th, 2017 Exam hours: 09:00 – 13:00 Permitted materials: None

The course teachers will visit the exam room at least once during the exam.

The exam text consists of problems 1-40 (multiple choice questions) to be answered by selecting true or false for each statement. If you think a statement could be either true or false, consider the most likely use/case. Problems 41-43 are answered by entering text. Problems 1-40 have a total weight of 80%, while problems 41-43 have a weight of 20%. Scoring in multiple choice questions Each problem has a variable number of true statements, but there is always at least

• ne true and one false statement for each problem. If you think a statement could be

either true or false, consider the most likely use/case. 0.5 point is given for each correctly marked statement. Further, an incorrectly marked statement or an unmarked statement(s) results in 0 point. The maximum score for a question is 2 points and the minimum is 0. It will be compensated through grade thresholds adjustments for the lack of negative points (to adjust for the opportunity to get a positive score by random answering).

Most likely use/case

• If you think a statement could be either true
• r false, consider the most likely use/case
• Example: “Evolutionary algorithms maintain a

population of candidate solutions”

– May be False for certain specific EAs – However, main focus in our class has been on EAs with a population

4

Multiple-choice Questions in Digital Exam

5

Digital exam: Text reply questions

6

We prefer answers on text problems in English language

but we would naturally not reduce the score due to spelling errors as long as the understanding seems to be correct like e.g.

• ” the bias node multiplied with it's respectful

weights is used to calculate the activation function in the first hidden layer.”

• Answer briefly and with structured and

formatted text.

7

Example: Ethical Recommendations for Robots

Structure and formatting

• With:
• Without

8

2018.11.21 9

INF3490/INF4490

Syllabus:

• Selected parts of the following books (details on course

web page):

– A.E. Eiben and J.E. Smith: Introduction to Evolutionary Computing, Second Edition (ISBN 978-3-662-44873-1). Springer. – S. Marsland: Machine learning: An Algorithmic Perspective. ISBN: 978-1466583283

– On-line papers (on the course web page).

• The lecture notes.

Supporting literature in Norwegian (not syllabus)

Jim Tørresen: hva er KUNSTIG INTELLIGENS Universitetsforlaget Nov 2013, ISBN: 9788215020211

10

Topics:

• Kunstig intelligens og

intelligente systemer

• Problemløsning med kunstig

intelligens

• Evolusjon, utvikling og læring
• Sansing og oppfatning
• Bevegelse og robotikk
• Hvor intelligente kan og bør

maskiner bli?

OPTIMIZATION AND SEARCH

Brief Summary

11

13

Search Landscapes

Some Optimization Methods

• 1. Exhaustive search
• 2. Greedy search and hill climbing
• 3. Simulated annealing
• 4. Gradient descent/ascent

– Not applicable for discrete optimization

Exploitation and Exploration

• Search methods should combine:

– Trying completely new solutions (like in exhaustive search) => Exploration – Trying to improve the current best solution by local search => Exploitation

15

16

17

EVOLUTIONARY ALGORITHMS

Brief Summary

18

The Problem with Hillclimbing

General scheme of EAs

20

Population Parents Parent selection Survivor selection Offspring Recombination (crossover) Mutation Intialization Termination

Genotype vs phenotype

Genotype Phenotype

21

Locus Loci

Representation and variation

• perators
• First stage of building an EA and most difficult one:

choose right representation for the problem

• Type of variation operators needed depends on

chosen representation

• Representations we have seen:

– Binary strings – Integers – Floating-point numbers – Permutations – Trees

22

23

24

Selection in EAs

• Selection can occur in two places:

– Parent selection (selects mating pairs) – Survivor selection (replaces population)

• Selection works on the population
• > selection operators are representation-

independent !

• Selection pressure: As selection pressure

increases, fitter solutions are more likely to survive, or be chosen as parents

25

Effect of Selection Pressure

• Low Pressure
• High Pressure

26

Selection

• Parent selection:

– Fitness Proportionate Selection – Rank-based Selection – Tournament Selection – Uniform Selection

• Survivor selection:

– Elitism – (µ,λ)-selection – (µ+λ)-selection

27

Summary: The standard EA variants

Name Representation Crossover Mutation Parent selection Survivor selection Specialty Genetic Algorithm

Usually fixed-length vector Any or none Any Any Any None

Evolution Strategies

Real-valued vector Discrete or intermediate recombination Gaussian Random draw Best N Strategy parameters

Evolutionary Programming

Real-valued vector None Gaussian One child each Tournament Strategy parameters

Genetic Programming

Tree Swap sub-tree Replace sub-tree Usually fitness proportional Generational replacement None

28

Performance Measures

• Performance measures (off-line)

– Efficiency (alg. speed, also called performance)

• Execution time
• Average no. of evaluations to solution (AES, i.e., number of

generated points in the search space)

– Effectiveness (solution quality, also called accuracy)

• Success rate (SR): % of runs finding a solution
• Mean best fitness at termination (MBF)
• “Working” measures (on-line)

– Population distribution (genotypic) – Fitness distribution (phenotypic) – Improvements per time unit or per genetic operator – …

29

30

31

Hybrid EAs

32

Multi-Objective Evolutionary Alogrithms

• Find a set of non-dominated solutions (approximation

set) following the criteria of: – convergence (as close as possible to the Pareto-

• ptimal front),

– diversity (spread, distribution)

33

Multi-Ojective EAs: Requirements

• 1. Way of assigning fitness and selecting

individuals,

– usually based on dominance

• 2. Preservation of a diverse set of points

– similarities to multi-modal problems

• 3. Remembering all the non-dominated

points you have seen

– usually using elitism or an archive

34

35

36

MACHINE LEARNING

Brief Summary

37

Characteristics of ML

• Learning from examples to analyze new data
• Generalization: Provide sensible outputs for

inputs not encountered during training

• Iterative learning process
• Types:

– Supervised Learning – Reinforcement Learning – Unsupervised Learning

38

Supervised learning

• Training data provided as pairs:
• The goal is to predict an “output” y from an “input x”:
• Output y for each input x is the “supervision” that is

given to the learning algorithm.

– Often obtained by manual annotation – Can be costly to do

• Most common examples

– Classification – Regression

( )

( )

( )

( )

( )

( )

{ }

1 1 2 2

, , , ,..., ,

P P

x f x x f x x f x

( )

= y f x

Neural Networks: McCulloch and Pitts Neurons

• Greatly simplified biological neurons.
• Sum the weighted inputs
• If total is greater than some threshold, neuron “fires”
• Otherwise does not

40

Axon Terminal Branches

• f Axon

Dendrites

Σ

x1 x2 w1 w2 wn xn x3 w3

The Perceptron Network

41

Inputs Outputs

42

Training a perceptron

Σ

x1 x2 xn

. . .

w1 w2 wn

a=Σi=1

n wi xi

1 if a ≥ q

y = 0 if a < q

y

{

inputs weights activation

• utput

q

What Can Perceptrons Represent?

2018.11.21 43

0,0 0,1 1,0 1,1 0,0 0,1 1,0 1,1

AND XOR

• Only linearly separable functions can be represented

by a perceptron

44

45

46

Minsky & Papert (1969) offered solution to XOR problem by combining perceptron unit responses using a second layer of units.

1 2 +1 +1 3

Solution for XOR : Add a Hidden Layer !!

Backpropagation

Rumelhart, Hinton and Williams (1986)

xk xi wki wjk

δj δk

yj

Backward step: propagate errors from

• utput to hidden layer

Forward step: Propagate activation from input to output layer

48

49

The Solution: Cross-Validation

To maximize generalization and avoid overfitting, split data into three sets:

• Training set: Train the model.
• Validation set: Judge the model’s generalization ability

during training.

• Test set: Judge the model’s generalization ability after

training.

50

51