+ All Categories
Home > Documents > Evolutionary Computation Evolutionary Learning...

Evolutionary Computation Evolutionary Learning...

Date post: 26-May-2020
Category:
Upload: others
View: 3 times
Download: 0 times
Share this document with a friend
22
Evolutionary Computation Evolutionary Computation Evolutionary Learning (I) Shan He School for Computational Science University of Birmingham Modules 02411 and 22313: EC
Transcript
Page 1: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Evolutionary ComputationEvolutionary Learning (I)

Shan He

School for Computational ScienceUniversity of Birmingham

Modules 02411 and 22313: EC

Page 2: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Outline

Outline of Topics

What is Evolutionary Learning?

Learning classifier systems

Evolutionary Artificial Neural Networks

Conclusion

Page 3: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

What is Evolutionary Learning?

What is Machine Learning?I Machine learning: How to construct programs that learn from

experience automatically [1]I More formal definition: A computer program is said to learn

from experience E with respect to some class of tasks T andperformance measure P, if its performance at tasks in T , asmeasured by P, improves with experience E

I The task can be:I Classification: learning how to label correctly new instances

from a domain based on experience E , e.g., a set of previouslylabeled instances

I Regression: estimating the relationships among variablesI Clustering: grouping a set of objects in such a way that

objects in the same group are more similar to each other thanto those in other groups variables.

I Association: where association between variables characterisingthe system is to be identified

[1] Machine Learning, Tom Mitchell, McGraw Hill, 1997.

Page 4: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

What is Evolutionary Learning?

Machine learning process

I Data preparation: obtain data / data checking/ datacleaning

I Feature selection: if you have high-dimensional data.

I Specify assumptions based on domain knowledge.

I Develop Model based on the assumptions.

I Specify loss function, or lost function e.g., the mean leastsquare error between the model output and the real data.

I Use algorithms to minimize the loss function based on thetrain data.

I Test the model using testing data

Page 5: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

What is Evolutionary Learning?

Case study 1: Australian Credit Card Assessment

I T : to assess applications to an Australian bank for a creditcard based on a number of attributes.

I E : 690 labelled application records, which consists ofI 2 classes: granted (44.5% of the instances) or denied (55.5%

of the instances)I 15 attributes (A1-A15): names and values have been changed

to meaningless symbols to protect confidentiality of the data.I Mixing-value inputs: there are 5 continuous, 4 binary and 5

nominalI A lot of missing value (999 in the file).

I P: Accuracy, e.g., percentage of correct classification

I You can download this file from my module web page anddetails can be found at UC Irvine Machine Learning Repository

Page 6: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

What is Evolutionary Learning?

Case study 2: Military Trauma survival prediction

Page 7: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

What is Evolutionary Learning?

More classification examples

I Classification example 1: handwriting recognitionI T : recognizing and classifying handwritten words within

imagesI E : a database of handwritten words with given classificationsI P: Accuracy, e.g., percentage of correct classification

I Classification example 2: spam detectionI T : classifying emails as spam or non-spam based on the

contentsI E : a database of email with given classificationsI P: Accuracy, e.g., percentage of correct classification

Page 8: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

What is Evolutionary Learning?

What is Evolutionary Learning?

I Evolutionary learning: machine learning methods based onevolutionary computation, also called Genetics-based MachineLearning

I Artificial evolution has been used in evolving the followinglearning machines:

I Rule-based systemsI Finite state machinesI Artificial Neural NetworksI Graphs

I The key issues here include representation and fitnessevaluation.

I Based on different representations, we have:I Learning classifier systems (this lecture)I Evolutionary Artificial Neural Networks (this lecture)I Genetic Programming (next lecture)

Page 9: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Learning classifier systems

What is Learning classifier systems (LCS)?I The oldest way of applying GAs to learning problemsI Basic idea: evolves condition-action (IF-THEN) rules for

classification.I Depending on the representation of individuals in a

population, there are Michigan Approach and PittsburghApproach

I Michigan Approach:I Each individual represents only a part of one solutionI One complete solution need to be represented by many

individualI Fitness is assigned to partial solutionsI Individuals not only compete for reproduction but also

cooperate with each otherI Very complex system: need to design an complex fitness

function to measure individuals’ contributions to the overallsolution

I The best set of individuals may not be the set of best (i.e.fittest) individuals

Page 10: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Learning classifier systems

Pittsburgh ApproachI Pittsburgh Approach: Each individual encodes one solution,

e.g., a set of rules to solve the problem.I How to represent rules as individuals (chromosomes)?I Take the Australian Credit Card Assessment problem as an

example:I Given the rules for classification, e.g.,

IF (c11) AND (c12) AND · · · AND (c1n) THEN Label = L1

IF (c21 ) AND (c22) AND · · · AND (c2n) THEN Label = L2...IF (ck1 ) AND (ck2 ) AND · · · AND (ckn) THEN Label= Lk

I Real-interval representation: cij is a real value and can be usedto represent lower/upper bound, e.g., if c11 = 10 and c12 = 20can mean attribute 10 ≤ A1 ≤ 20

I One simple representation: encode them in a single individual(chromosome)c11, c12, · · · , c1n, L1, c21, c22, · · · , c2n, L2, · · · , ck1, ck2, · · · , ckn, Lk

Page 11: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Learning classifier systems

Discussion about Pittsburgh Approach (I)I Fixed-length representation is used in previous example.I In the fixed-length representation, we can allow dummy rules

with empty conditions and actions, e.g., when the upperbound is less than the lower.

I We can also use variable chromosome length if necessary.I Conflict: two rules whose conditions are the same, but the

labels are different, we need to decide which rule is used inthese cases

I Match process: The order of the rules may imply the order ofrule firing (activation), e.g., the conditions of rules 2, 3 and 7are the same, rule 2 will be used

I Domain knowledge can be incorporated into LCS by biasinitialisation, e.g., non-uniform distribution

I If we know nothing about the problem, we can use uniformrandom initialisation

Page 12: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Learning classifier systems

Discussion about Pittsburgh Approach (II)

I Variation operators in standard GA, e.g., one-point crossoverand uniform mutation can be used in Pittsburgh Approach

I Fitness calculation essentially evaluate how well eachindividual (i.e., a rule-based system) performs.

I It is essential that fitness evaluation emphasises maximisinggeneralisation, not simple memorisation of its experience e.g.,a training set.

I Generalisation in machine learning: generalize from itsexperience, which means the trained learning machine canperform accurately on new, unseen examples/tasks

I Essentially to avoid overfitting, e.g., “describes random erroror noise instead of the underlying relationship”

Page 13: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Learning classifier systems

Overfitting

Page 14: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Learning classifier systems

Quiz: which fitness funtion is better?

I 1. Based on the training error, e.g., the mean square error

I 2. Based on the error on a separate validation set

I 3. Based on the training error and complexity

Page 15: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Learning classifier systems

Pittsburgh Approach Fitness Calculation

NumExamples := 0;CorrectExamples := 0For each training example in training set

NumExamples := NumExamples + 1Determine the first rule that matches the training exampleIf output of the rule is as same as the true label:

CorrectExamples := CorrectExamples + 1;Output CorrectExamples/NumExamples - RuleComplexityMeasure

Page 16: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Evolutionary Artificial Neural Networks

What are Artificial Neural Networks (ANNs)?

Hidden LayerInput Output

I ANN: Mathematical model or computational model inspiredby biological neural networks.

I Consists of an interconnected group of artificial neurons

Page 17: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Evolutionary Artificial Neural Networks

What are Artificial Neural Networks (ANNs)?I Non-linear statistical data modeling tools:

I Model complex relationships between inputs and outputs;I Discover patterns in data.

I Can be used for classification, association, regression andclustering.

I Gradient-based learning algorithms (such as BackPropagation) are often used

I Drawbacks:I local search: they often get trapped in a poor local minimum;I they are sensitive to initial conditions;I they require the error function to be differentiable;I they can be very slow;I they may not help improve generalisation

I MATLAB Neural Network Toolbox (Click for more detailedtutorial)

Page 18: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Evolutionary Artificial Neural Networks

What are Evolutionary ANNs?I Also called Neuroevolution, based means using evolutionary

algorithms to train artificial neural networksI Advantage: more versatile than standard ANNs, e.g., we only

need to know a measure of a network’s performance at a taskI Two types of EANNs:

I Conventional neuroevolution: only evolve the connectionweights of a ANN of a fixed network topology

I Topology & Weight neuroevolution: evolve both the topologyof a network and its weights

I Essentially two ways of representing a ANN:I Conventional neuroevolution: a individual (chromosome)

represent the connection weights of the ANN, e.g., each geneis a connection weight between two nodes

I Topology & Weight neuroevolution: need to decode from achromosome to represent the topology and weights of anetwork

Page 19: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Evolutionary Artificial Neural Networks

Marker based encoding

I A individual (chromosome) is a string of integers representingboth the topology and connection weights of a network

I Two special values in a chromosome, called start marker andend marker, segment the chromosome into segments,

I Each segment defines a neuron.

I Each neuron has a key (id) which is defined by the value nextto the start marker

I All other values are used in pairs (ki , wi ) to define theconnection destination (id of another neuron) and theconnection weight.

Page 20: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Evolutionary Artificial Neural Networks

Marker based encoding

Page 21: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Conclusion

Conclusion

I Evolutionary rule-based systems can learn and adaptcontinuously, unlike classical expert systems or rule-basedsystems that are fixed after construction.

I Evolutionary ANNs overcome a lot of drawbacks of standardANNs but it is slow

I Chromosome representation of rules or ANNs has animportant impact on the success of evolutionary learning.

I Fitness evaluation needs to consider generalisation carefully.

I Domain knowledge and heuristics can be used effectively inevolutionary learning to produce hybrid learning systems thatare both effective and efficient.

Page 22: Evolutionary Computation Evolutionary Learning (I)szh/teaching/ec/Lecture16_EvolutionaryLearning.pdflearning machines: I Rule-based systems I Finite state machines I Arti cial Neural

Evolutionary Computation

Conclusion

Further reading

I T. Kovacs. Genetics-based Machine Learning, Handbook ofNatural Computing: Theory, Experiments, and Applications.Springer Verlag, 2012. Click here for an online version

I X. Yao, Evolving articial neural networks, Proceedings of theIEEE, 87(9):1423-1447, September 1999.


Recommended