Studies in Machine Learning

1
Studies in Machine Learning
159.734
Dr. Napoleon H. Reyes, Ph.D.
Computer Science
Institute of Information and Mathematical Sciences
Rm. 2.56 QA, IIMS, Albany Campus
Email n.h.reyes_at_massey.ac.nz Tel. No. 64 9
4140800 x 9512 Fax No. 64 9 441
8181 Consultation Hours After every lecture for
1 hr Lectures Mon. 11pm (IIMS Lab 1) Fri., 10am
(IIMS Lab 1)
2
Student Responsibility
159.734
Note
If a student cannot attend lectures/tutorials it
is the students responsibility to find out what
was discussed in lectures / tutorials (possible
changes to assignments, questions answers).
3
Topics for Discussion
159.734
Pre-requisites
Course Overview
Learning Outcomes
Texts and Course Material
Assessment
Course Schedule
4
Pre-requisites
159.734
Students are expected to
Have taken up Programming Fundamentals
5
Course Overview
159.734
Calendar Prescription
Theories of Machine Learning
Experiments and Applications of ML Algorithms
Recent Advances in ML
Problem-based learning
WEKA Machine Learning Tool and other ML Libraries
will be used for investigating algorithms
Hand-simulation of algorithms
C implementation of some selected algorithms
6
Learning Outcomes
159.734
On successful completion of the course, the
students should be able to
Understand and describe the main machine learning
algorithms used in computing.
Design and program computers to store and
manipulate knowledge.
Describe, and in some cases program, the special
algorithms that can be used to perform
intelligent tasks such as learning from examples,
learning from rules, predicting outcomes,
recognizing characters and navigating robots by
computers.
Identify the advantages and disadvantages of
applying various AI algorithms in solving
real-world problems.
7
Texts and Course Material
159.734
References
Machine Learning Notes from MITOpenCourseWare
Tutorials from http//www.cs.waikato.ac.nz/ml/
weka/
Data Mining Practical Machine Learning Tools and
Techniques (Second Edition) by Ian H. Witten,
Eibe Frank
Neural Network and Fuzzy Logic Applications in
C/C by Stephen T. Welstead (John Wiley Sons
1994)
Artificial Intelligence A Modern Approach, Second
Edition, (Prentice Hall 2003) by Russell, S. and
Norvig, P.
8
Assessment
159.734
2 assignments (30) 40
Final Exam (3 hours) (70) 60

• In order to pass the course you must have (both
  of them).
• The exam mark at least 45 from max.
• The final mark greater than or equal to 50.
• All assignments will be submitted electronically.

9
Assessment
159.734
Program solutions that do not compile or do not
run in our laboratories get 0 marks.
Late assignments will be penalized
Assignments may be completed in groups all
members of the group should be named in the
source file of each assignment they contributed.
10
Assessment
159.734
Each group member will receive the same grade.
Students in a team have the authority (in
consultation with the lecturer) to "expel" any
member that does not meet obligations .
The collaboration is limited only to members
within each group.
It is a student responsibility to check their
assignment marks and notify in writing any errors
they might find no later than 10 days after the
day the marks were made available.
11
Course Schedule
159.734
1. Introduction

• Learning
• Kinds of Learning
• - Supervised Learning
• - Clustering
• - Reinforcement
• Learning a Function
• - Memory
• - Averaging
• - Generalization
• Overview of the Different Learning Algorithms
• - Nearest Neighbor
• - Decision Trees
• - Neural Networks

12
Course Schedule
159.734
2. Supervised Learning

• Learning Conjunctions
• Learning Disjunctive Normal Form (DNF)
• Cross Validation
• Learning Curves
• Noisy Data
• Bias vs. Variance
• Learning Decision Trees
• Trees vs. DNF

13
Course Schedule
159.734
3. Naïve Bayes

• Learning Algorithm
• Prediction Algorithm
• Laplace Correction
• Hypothesis Space
• Bayes Rule

14
Course Schedule
159.734
4. Nearest Neighbor

• Feature Spaces
• Scaling
• Hypothesis of Nearest Neighbors
• Time and Space Requirements
• k-Nearest Neighbor
• Curse of Dimensionality

15
Course Schedule
159.734
5. Decision Trees

• Numerical Attributes
• Decision Trees with Continuous Features
• Example Domains

16
Course Schedule
159.734
6. Regression

• Numerical Attributes
• Decision Trees with Continuous Features
• Example Domains

17
Course Schedule
159.734
7 8 Linear Separators

• Linear Hypothesis Class
• Hyperplane
• Linear Classifier
• Perceptron Algorithm
• Gradient Ascent/Descent
• Perceptron Training
• Characterising a separator and margin
• Constrained Optimisation
• Dual Lagrangian

18
Course Schedule
159.734
9. Support Vector Machines

• SVM Classifier
• Key Points on SVM Training and Classification
• Kernel Functions
• - Linear Kernel
• - Polynomial Kernel
• - Radial Basis Kernel
• Cross-Validation Error

19
Course Schedule
159.734
10. Artificial Neural Networks

• Single Perceptron
• Multi-Layer Perceptron
• Generalized Delta Rule
• Back Propagation
• Training Neural Nets
• Momentum
• Input/Output Representation

20
Course Schedule
159.734
11 12. Features

• Feature Selection
• Feature Ranking
• Correlations in Data
• Subset Selection
• Forward Selection
• Backward Elimination
• Clustering
• - K-Means Clustering
• Principal Components Analysis
• Validating a Classifier
• Tunable Classifiers
• ROC Curves
• Review for Finals

21
Course Schedule
159.734
Tasks

• Send your email address to n.h.reyes_at_massey.ac.nz
  to receive announcements, tips, etc. during the
  duration of the semester
• Download WEKA 3.4.12 latest stable version
• Check out MIT OpenCourseWare on Artificial
  Intelligence, Machine Learning
• Check out UCI Machine Learning Repository
• Download the Letter Recognition data set from
  UCI
• Try to get hold of some of the references for
  the paper

22
Course Schedule
159.734
Demo

• Neural Network XOR Problem
• WEKA
• Self-Organizing Maps
• Genetic Algorithm
• Character Recognition

23
Fuzzy System
Inverted Pendulum Problem / Broom-Balancing
Problem
Input x, v, theta, angular velocity
Output Force, direction
24
Fuzzy System
Inverted Pendulum Problem
If the broom angle is too big or changing too
quickly, then regardless of the location of the
cart on the cart path, push the cart towards the
direction it is leaning to.
If the cart is too near the end of the path, then
regardless of the state of the broom angle push
the cart towards the other end.
Input x, v, theta, angular velocity
Output Force, direction
25
Cascade of Fuzzy Systems
Multiple Fuzzy Systems employ the various robot
behaviours
Path planning Layer The A Algorithm
Path Planning Layer
Next Waypoint
Fuzzy System 1 Target Pursuit
Fuzzy System 1
Target Pursuit
Adjusted Angle
Central Control
Fuzzy System 2 Speed Control for Target Pursuit
Fuzzy System 2
Adjusted Speed
ObstacleDistance lt MaxDistanceTolerance and
closer than Target
N
Y
Fuzzy System 3 Obstacle Avoidance
Fuzzy System 3
Obstacle Avoidance
Adjusted Angle
Fuzzy System 4
Fuzzy System 4 Speed Control for Obstacle
Avoidance
Adjusted Speed
Actuators
26
Cascade of Fuzzy Systems
Obstacle Avoidance, Target Pursuit, Opponent
Evasion
Input Multiple Obstacles x, y,
angle Targets x, y, angle
Output Robot angle, speed
27
Simulations
Hybrid Fuzzy A
Robot Navigation System
3-D Hybrid Fuzzy A Navigation System
Cascade of Fuzzy Systems
28
Fuzzy A
Input Obstacles x, y, angle Targets x, y,
angle
Output Robot angle, speed
29
A Simple Kohonen Network
Neural Network Architecture with Unsupervised
Learning
Lattice
4x4
Node
Weight Vectors
Input Nodes
Input Vector
30
SOM for Color Clustering
Unsupervised learning
Reduces dimensionality of information
Clustering of data
Topological relationship between data is
maintained
Input 3D , Output 2D
Vector quantisation
31
Genetic Algorithm
Optimised Search

• GAs are biologically inspired class of algorithms
  that can be applied to, among other things, the
  optimization of nonlinear multimodal functions.
• Solves problems in the same way that nature
  solves the problem of adapting living organisms
  to the harsh realities of life in a hostile
  world evolution.

32
Genetic Algorithm
Optimised Search

• Potential solutions to a problem are investigated
  through the application of an evolution process,
  allowing chromosomes to mate and mutate. In
  the end, the best offspring is selected.

33
A Simple GA Example

• Function to evaluate
• coeff chosen to normalize the x parameter when
  a bit string of length lchrom 30 is chosen.
• Since the x value has been normalized, the max.
  value of the function will be
• when for the case
  when lchrom30

Fitness Function or Objective Function
34
Test Problem Characteristics

• With a string length30, the search space is much
  larger, and random walk or enumeration should not
  be so profitable.
• There are 2301.07(1010) points. With over 1.07
  billion points in the space, one-at-a-time
  methods are unlikely to do very much very
  quickly. Also, only 1.05 percent of the points
  have a value greater than 0.9.

35
?
END