Artificial Intelligence Lecture 15 Introduction to Machine Learning
Overview Machine Learning  ID3 Decision Tree Algorithm Discussions
Machine Learning Supervised Learning Training examples consist of pairs of input vectors, and desired outputs  Unsupervised Learning Training examples do not contain hints about correct outputs Usually used to identify unusual structures in data
Inductive Learning Basics Inferring a boolean/real-valued function from training examples A training example is a pair of (x, f(x)) x is the input f(x) is the output of the function applied to x
Hypothesis Any function that approximates the given set of examples Bias : preference for one hypothesis beyond mere consistency (a) (b) (c) (d)
Hypothesis Space A set of all hypotheses consistent with data denoted by H:{H 1 , H 2 , …, H n } Inductive learning is searching for a good hypothesis in the hypothesis space Occam’s razor: prefer the simplest hypothesis consistent with data Inductive Learning Assumption : Any hypothesis found to  approximate the target  function well over a  sufficiently large  set of  training examples  will also  approximate the target  function well over  other unobserved examples .
ID3 Decision Tree Algorithm
Training Examples
Decision Tree for  PlayTennis ? ? PlayTennis Wind Humidity Temperature Outlook Weak High Hot Overcast Weak High Hot Sunny
Decision Tree Representation Decision Tree Each internal node tests an attribute Each branch takes an attribute value Each leaf node predict a class label Disjunction of conjunctions of a set of attribute values
Entropy
Information Gain
ID3 Algorithm
Training Examples
Selecting the Best Attribute
 
Hypothesis Space Search by ID3
Discussions
Hypothesis Space Search by ID3 Hypothesis space is complete Target function surely in there … Outputs a single hypothesis (Which one?) Cannot determine how many alternatives No back tracking Local minima … Use statistical properties of all training data at each step in search Robust to noisy data
Inductive Bias in ID3 Note the hypothesis space H is the power set of instances X. Unbiased? Not really. Preference for short trees, and for those with high information gain attributes near the root Bias is a preference for some hypotheses, rather than a restriction of hypothesis space H, e.g. target concept is not in H.
Summary Machine Learning ID3 Decision Tree Algorithm Discussions

Introduction to Machine Learning

  • 1.
    Artificial Intelligence Lecture15 Introduction to Machine Learning
  • 2.
    Overview Machine Learning ID3 Decision Tree Algorithm Discussions
  • 3.
    Machine Learning SupervisedLearning Training examples consist of pairs of input vectors, and desired outputs Unsupervised Learning Training examples do not contain hints about correct outputs Usually used to identify unusual structures in data
  • 4.
    Inductive Learning BasicsInferring a boolean/real-valued function from training examples A training example is a pair of (x, f(x)) x is the input f(x) is the output of the function applied to x
  • 5.
    Hypothesis Any functionthat approximates the given set of examples Bias : preference for one hypothesis beyond mere consistency (a) (b) (c) (d)
  • 6.
    Hypothesis Space Aset of all hypotheses consistent with data denoted by H:{H 1 , H 2 , …, H n } Inductive learning is searching for a good hypothesis in the hypothesis space Occam’s razor: prefer the simplest hypothesis consistent with data Inductive Learning Assumption : Any hypothesis found to approximate the target function well over a sufficiently large set of training examples will also approximate the target function well over other unobserved examples .
  • 7.
  • 8.
  • 9.
    Decision Tree for PlayTennis ? ? PlayTennis Wind Humidity Temperature Outlook Weak High Hot Overcast Weak High Hot Sunny
  • 10.
    Decision Tree RepresentationDecision Tree Each internal node tests an attribute Each branch takes an attribute value Each leaf node predict a class label Disjunction of conjunctions of a set of attribute values
  • 11.
  • 12.
  • 13.
  • 14.
  • 15.
  • 16.
  • 17.
  • 18.
  • 19.
    Hypothesis Space Searchby ID3 Hypothesis space is complete Target function surely in there … Outputs a single hypothesis (Which one?) Cannot determine how many alternatives No back tracking Local minima … Use statistical properties of all training data at each step in search Robust to noisy data
  • 20.
    Inductive Bias inID3 Note the hypothesis space H is the power set of instances X. Unbiased? Not really. Preference for short trees, and for those with high information gain attributes near the root Bias is a preference for some hypotheses, rather than a restriction of hypothesis space H, e.g. target concept is not in H.
  • 21.
    Summary Machine LearningID3 Decision Tree Algorithm Discussions