chapter 10 machine learning: symbol-based

CSC411CSC411 Artificial IntelligenceArtificial Intelligence 11

Chapter 10

Machine Learning: Symbol-Based

ContentsContentsA FrameworkVersion Space SearchID3: Decision Tree


Machine LearningMachine LearningAI systems grow from a minimal amount of AI systems grow from a minimal amount of knowledge by learningknowledge by learningHerbert Simon (1983):Herbert Simon (1983):– Any change in a system that allows it to Any change in a system that allows it to

perform better the second time on repetition of perform better the second time on repetition of the same task or on another task drawn from the same task or on another task drawn from the same populationthe same population

Machine learning issues:Machine learning issues:– Generalization from experienceGeneralization from experience

InductionInductionInductive biasesInductive biases

– Performance change: improve or degradePerformance change: improve or degrade


Machine Learning CategoriesMachine Learning CategoriesSymbol-based learningSymbol-based learning– Inductive learning -- learning by examplesInductive learning -- learning by examples– Supervised learning/unsupervised learningSupervised learning/unsupervised learning

Concept learning –- classificationConcept learning –- classification

Concept formation -- clusteringConcept formation -- clustering

– Explanation-based learningExplanation-based learning– Reinforcement learningReinforcement learning

Neural/connectionist networksNeural/connectionist networks

Genetic/evolutionary learning Genetic/evolutionary learning


A general model of the learning process


Learning ComponentsLearning ComponentsData and goals of learning taskData and goals of learning task– What are given – training instancesWhat are given – training instances– What are expectedWhat are expected

Knowledge representationKnowledge representation– Logic expressionsLogic expressions– Decision treesDecision trees– RulesRules

OperationsOperations– Generalization/specializationGeneralization/specialization– Heuristic rules Heuristic rules – Weight adjustsWeight adjusts

Concept spaceConcept space– Search space: representation, formatSearch space: representation, format

Heuristic searchHeuristic search– Search control in the concept spaceSearch control in the concept space


Learning By ExamplesLearning By ExamplesPatrick Winston (1975)Patrick Winston (1975)– Given a set of positive and a set of negative Given a set of positive and a set of negative

examplesexamples– Find a concept representationFind a concept representation– Semantic network representationSemantic network representation

ExampleExample– Learn a general definition of structural Learn a general definition of structural

concept, say “concept, say “archarch””– Positive examples: examples of Positive examples: examples of archarch

What an arch looks like, to define the archWhat an arch looks like, to define the arch

– Negative examples: near missesNegative examples: near missesWhat an arch doesn’t look like, to avoid the over-What an arch doesn’t look like, to avoid the over-coverage of archcoverage of arch


Examples and near misses for the concept “arch.”


Generalization of descriptions to include multiple examples.


Generalization of descriptions to include multiple examples (cont’d)


Specialization of a description to exclude a near miss. In c we add constraints to a so that it can’t match with b.


Version Space SearchVersion Space SearchInductive learning as search through a Inductive learning as search through a concept spaceconcept spaceGeneralization imposes an ordering on the Generalization imposes an ordering on the concepts in the space and uses the concepts in the space and uses the ordering to guide the searchordering to guide the searchGeneralizationGeneralization– PrinciplesPrinciples

Extend the coverage of instances Extend the coverage of instances Shorten/shrink the constrainsShorten/shrink the constrains

– Operations Operations Replacing constant with variablesReplacing constant with variablesDropping conditions from a conjunctive expressionDropping conditions from a conjunctive expressionAdding a disjunct to an expressionAdding a disjunct to an expressionReplacing a concept with one of its parent conceptsReplacing a concept with one of its parent concepts


A concept space: • Initial state obj(X, Y, Z) might cover all instances: too general• As more instances are added, X, Y, Z will be constrained


Version Space Search AlgorithmsVersion Space Search AlgorithmsCharacteristics of these algorithmsCharacteristics of these algorithms

– Data-drivenData-drivenPositive examples to generalize the conceptPositive examples to generalize the conceptNegative examples to constrain the concept (avoid Negative examples to constrain the concept (avoid overgeneralization)overgeneralization)

– Procedure:Procedure:Starting from whole space Starting from whole space Reducing the size of the space as more examples included Reducing the size of the space as more examples included Finding regularities (rules) in the training dataFinding regularities (rules) in the training data

– Generalization on these regularities (rules)Generalization on these regularities (rules)

Three algorithmsThree algorithms– Reducing the size of the version space in a Reducing the size of the version space in a specific to specific to

generalgeneral direction direction– Reducing the size of the version space in a Reducing the size of the version space in a general to general to

specificspecific direction direction– Combination of above: Combination of above: candidate elimination algorithmcandidate elimination algorithm


The role of negative examples in preventing overgeneralization by forcing the learner to specialize concepts in order to exclude negative examples

Negative ExamplesNegative Examples


Specific to General SearchSpecific to General SearchMaintains a set S of candidate concepts, Maintains a set S of candidate concepts, the maximally specific generalizations the maximally specific generalizations from the training instancesfrom the training instances

A concept c is maximally specific if it A concept c is maximally specific if it – covers all positive examples, non of the covers all positive examples, non of the

negative examples, and negative examples, and – for any other concept c’ that covers the for any other concept c’ that covers the

positive examples, c≤c’positive examples, c≤c’

The algorithm usesThe algorithm uses– Positive examples to generalize the candidate Positive examples to generalize the candidate

conceptsconcepts– Negative example to avoid overgeneralizationNegative example to avoid overgeneralization


For hypothesis set S:Specific to General Search AlgorithmSpecific to General Search Algorithm


Specific to general search of the version space learning the concept “ball.”


General to Specific SearchGeneral to Specific SearchMaintains a set G of maximally general Maintains a set G of maximally general concepts concepts A concept c is maximally general if it A concept c is maximally general if it – covers non of the negative training examples, covers non of the negative training examples,

and and – for any other concept c’ that covers no for any other concept c’ that covers no

negative training examples, cnegative training examples, cc’c’

The algorithm usesThe algorithm uses– negative examples to specialize the candidate negative examples to specialize the candidate

concepts concepts – Positive examples to eliminate Positive examples to eliminate

overspecializationoverspecialization


General to Specific Search AlgorithmGeneral to Specific Search Algorithm


General to specific search of the version space learning the concept “ball.”


Candidate Elimination AlgorithmCandidate Elimination Algorithm

Combination of above two algorithms into a bi-direction searchMaintains two sets of candidate concepts– G, the set of maximally general candidates– S, the set of maximally specific candidates

The algorithm specializes G and generalizes S until they converge on the target concept.


Candidate Elimination AlgorithmCandidate Elimination Algorithm


The candidate elimination algorithm learning the concept “red ball.”


Converging boundaries of the G and S sets in the candidate elimination algorithm.


Decision TreesDecision TreesLearning algorithms of inducing concepts Learning algorithms of inducing concepts from examplesfrom examples

CharacteristicsCharacteristics– A tree structure to represent the concept, A tree structure to represent the concept,

equivalent to a set of rulesequivalent to a set of rules– Entropy and information gain as heuristics for Entropy and information gain as heuristics for

selecting candidate conceptsselecting candidate concepts– Handling noise dataHandling noise data– Classification – supervised learningClassification – supervised learning

Typical systems: ID3, C4.5, C5.0Typical systems: ID3, C4.5, C5.0


Data from credit history of loan applications


A decision tree for credit risk assessment.


A simplified decision tree for credit risk assessment.


The induction algorithm begins with a sample of correctly classified members of the target categories.

Decision Tree Construction AlgorithmDecision Tree Construction Algorithm


A partially constructed decision tree.

Another partially constructed decision tree.

chapter 10 machine learning: symbol-based

Documents

concept arch

parent conceptsa concept

archpositive examples

examples of archwhat

conceptnegative examples

archnegative examples

concept spacelearning

concept spacegeneralization