
	2D1431 Machine Learning


requirements	teacher	schedule	newsgroup	labs	lectures	course pages

Machine Learning

Frank Hoffmann

News:

20/12/02: Exam solution available
28/11/02: New lab description for lab4 available
21/11/02: New lab description for lab3 available
11/11/02: New lab description for lab2 available
11/11/02: The lecture on Friday 22/11/02 13-15 has been moved to Monday 18/11/02 15-17 in D3

Course homepage: 2D1431

Course description 2D1431: swedish and english

Start: VT 2002, Period 2:
first lecture: 23-10-2002 13.00 E3

Newsgroup:The course newsgroup nada.kurser.mi serves as a forum for announcements, technical questions, help and exchange of ideas related to the course and the labs. Please check it regularly for new messages.

Schema: schedule

Registration:
Register in RES and subscribe to the course environment with the UNIX commands:
res checkin mi02
course join mi02
For the labs you can reserve an examination (redovisning) time in advance using the RES system with the commands:
module add resultat
bok new mi02
For more information on res and bok use the commands res help and bok help

Teacher:
lecturer: Frank Hoffmann hoffmann@nada.kth.se
I am located at CVAP, here are directions on how to get to CVAP. Please call at 790-6271 in advance to make sure that I am in my office.

lab assistents:

Frank Hoffmann hoffmann@nada.kth.se
Mikael Huss hussm
Martin Rehn rehn
Christopher Johansson cjo

Synopsis:
Machine learning is concerned with computer programs that automatically improve their performance with past experiences. Machine learning draws inspiration from many fields, artificial intelligence, statistics, information theory, biology and control theory.

The objective of this course is to introduce the student to the basic theory and algorithms that form the foundation of machine learning. The course will cover the following topics

introduction
concept learning
decision tree learning
artificial neural networks
instance based learning
evolutionary algorithms
reinforcement learning
fuzzy logic
Bayesian learning
computational learning theory
machine learning in robotics

Exam:
Exam 14/12/02 solution
Saturday: 14/12/02 8-13, L21-22, L43, L51
Since the course is taught for the first time there are no previous exams. However, to get an idea on what type of questions to expect take a look at the exams in the course 2D1432 Artificial Neural Networks.

Requirements and Grading:
The course is largely self-contained and has no prerequisites other than basic knowledge of computer science and programming experience. To earn credits, a student has to complete the four mandatory lab assignments and to pass a written exam. The labs have to be presented at the specified dates in order to get bonus points. Handing in labs after the scheduled redovisning date results in loss of bonus points to improve your grade.
The grade (betyg) for the course is determined by the outcome of the exam. The maximum number of points in the exam is 40 points, the grades are

Students with a grade 3,4,5 in the exam can improve the grade for the course by presenting lab assignments in time. For each lab presented in time you receive 1.5 bonus points, which means that if you present all four labs in time, you will improve your overall grade by 1. In order to pass the course you have to achieve at least a grade 3 in the exam. Example: a student achieves 25 points in the exam and presents all four labs in time (4x1.5=6 bonus points), the final grade for the course is 4 (25p+6p=31p).

Course literature:
required:

T. Mitchell, Machine Learning;, McGrawHill, 1997, hardcover ISBN 0-07-042807-7, paperback ISBN 0-07-115467-1
The hard copy costs $107.50 if you order from the publisher. However, you can purchase the less expensive paperback, which for example at Amazon UK was available for £ 34 (new) and £ 10.20 (used).
selected papers on particular topics

recommended:

Reinforcement Learning: An Introduction (online textbook) by Richard Sutton

Labs:
The course will contain four mandatory labs. Students prepare the solutions to the lab assignments prior to the scheduled lab. During the lab you present your program and answers to the question to the assistent to obtain credits for the lab. Labs can be presented in groups of two, however both students need to fully understand the entire solution and answers. It is also assumed that you complete the assignment on your own and do not use parts of someone else's code or solution. Violation of these rules may result in failing the entire course.

You can reserve a time slot for examination (redovisning) in advance with the command:
bok new mi02
If you need to cancel a reservation use the command:
bok remove mi02

After you presented your lab to the assistent let him sign the lab receipt (labkvitto).

The labs will be programmed in Matlab. It is assumed that you are familiar with the basic features of Matlab. If not you should take a look at the following tutorials and documentation.

Matlab Online Documentation at NADA
Matlab Help Desk at NADA
Matlab Primer (PDF) by Kermit Sigmon
A Practical Introduction to Matlab by Mark S. Gockenbach
Matlab at Google

The labs will use routines for pattern classification from the Matlab Toolbox for Pattern Recognition at Delft University

Lab1: Concept Learning and Decision Trees : 14-11-2002
lab description: lab1.pdf
Matlab files and datasets: lab1.tar
If you work under Windows and Linux you need the M-file oracle.m instead of the MEX-file oracle.mexsol
Lab2: Bayes Classifier, Bagging and Boosting : 21-11-2002
lab description: lab2.pdf
Matlab files and datasets: lab2.tar
literature & links:
- Boosting.org
- R.E. Schapire. A brief introduction to boosting. In Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence, 1999.
- J.R. Quinlan. Bagging, boosting, and C4.5 In Proceedings of the Thirteenth National Conference on Artificial Intelligence and the Eighth Innovative Applications of Artificial Intelligence Conference, pages 725-730, Menlo Park, August4-8 1996, AAAI Press / MIT Press
- E. Bauer, R. Kohavi An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants, 1999, Machine Learning
Lab3: Instance Based Learning, Artificial Neural Networks : 28-11-2002
lab description: lab3.pdf
Matlab files and datasets: lab3.tar
literature & links:
- C. Atkeson, A. Moore, S. Schaal , Locally Weighted Learning, Artificial Intelligence Review, 11:11-73, 1997
- Ron Kohavi, A Study of Cross-Validation and Bootstrap for Accuracy Estimation and Model Selection, 1995, IJCAI-95
- George John, Ron Kohavi, Karl Pfleger Irrelevant Features and the Subset Selection Problem, 1994, Proceedings of the Eleventh International Conference on Machine Learning
Lab4: Reinforcement Learning : 5-12-2002
lab description: lab4.pdf
Matlab files: lab4.tar
literature & links:
- Reinforcement Learning: A Survey by Kaelbling, L.P., Littman, M.L., and Moore, A.W, 1996.
- Reinforcement Learning: An Introduction by Richard Sutton
  in particular chapters on

Lectures:

Lecture 1: Introduction to Machine Learning : 23-10-2002 & 24-10-2002
Slides:

Lecture slides (PDF, 4 slides per page)
Lecture slides (PDF, 1 slide per page)
Lecture slides (Powerpoint)
Links:

Learning to play backgammon
Golem project
Homework:

read Mitchell : chapter 1
read Gerald Tesauro : TD-Gammon, a self-teaching backgammon program, achieves master-level play 1994,
(local)
Additional reading:

Artifical Intelligence : A Modern Approach, Stuart Russell & Peter Norvig: chapter 5 - Game Playing
Lecture 2: Concept Learning : 30-10-2002
Slides:
Homework:
- read Mitchell : chapter 2
Lecture 3: Decision Tree Learning : 31-10-2002
Slides:
Homework:
- read Mitchell : chapter 3
Additional reading:
- Sreerama K. Murthy, Automatic Construction of Decision Trees from Data: A Multi-Disciplinary Survey", Data Mining and Knowledge Discovery, 2 (4), 345-389, 1998, available from Kluwer online
Lecture 4: Artificial Neural Networks : 6-11-2002 & 8-11-2002
Slides:
Homework:
- read Mitchell : chapter 4
Additional reading:
- Neural Networks - A Comprehensive Foundation, Simon Haykin, Prentice-Hall, 1999, Matlab code
- Neural Networks for Pattern Recognition, C.M. Bishop, Oxford University Press, 1996
- Neural Network Design, Martin T. Hagan, Howard B. Demuth, Mark H.Beale, PWS, 1995, Matlab demo
Software:
Links:
- European Network of Excellence on Neural Networks
- Neural Networks at Google
Lecture 5: Bayesian Learning : 13-11-2002 & 15-11-2002
Slides:
Homework:
- read Mitchell : chapter 6
Additional reading:
- A Tutorial on Learning with Bayesian Networks, David Heckerman , 1996
- Artifical Intelligence : A Modern Approach, Stuart Russell & Peter Norvig: chapters 14+15 - Uncertainty + Probabilistic Reasong Systems
- Pattern classification, Richard O. Duda, Peter E. Hart, David G. Stork: chapters 2 + 3 - Bayesian Decison Theory + Maximum Likelihood and Bayesian Parameter Estimation
Software:
- BayesNet Toolbox for Matlab
Lecture 6: Bagging & Boosting : 15-11-2002
Slides:
Homework:
- read (mandatory) : R.E. Schapire. A brief introduction to boosting. In Proceedings of the Sixteenth International Joint Conference on Artificial Intelligence, 1999.
Additional reading:
- Pattern classification, Richard O. Duda, Peter E. Hart, David G. Stork: chapter 9 - Algorithm Independent Machine Learning
- J.R. Quinlan. Bagging, boosting, and C4.5 In Proceedings of the Thirteenth National Conference on Artificial Intelligence and the Eighth Innovative Applications of Artificial Intelligence Conference, pages 725-730, Menlo Park, August4-8 1996, AAAI Press / MIT Press
- E. Bauer, R. Kohavi An Empirical Comparison of Voting Classification Algorithms: Bagging, Boosting, and Variants, 1999, Machine Learning
- George John, Ron Kohavi, Karl Pfleger Irrelevant Features and the Subset Selection Problem, 1994, Proceedings of the Eleventh International Conference on Machine Learning
Links:
- Boosting.org
Lecture 7: Instance Based Learning : 18-11-2002
Slides:
Homework:
- read Mitchell : chapter 8
Additional reading:
- Pattern classification, Richard O. Duda, Peter E. Hart, David G. Stork: chapter 4 - Non-parametric techniques
- C. Atkeson, A. Moore, S. Schaal , Locally Weighted Learning, Artificial Intelligence Review, 11:11-73, 1997
- C. Atkeson, A. Moore, S. Schaal , Locally Weighted Learning for Control, Artificial Intelligence Review, 11:75-113, 1997
Lecture 8: Reinforcement Learning : 20-11-2002 & 27-11-2002
Slides:
Homework:
- read Mitchell : chapter 13
- read at least first four chapters (pages 1-19) : Reinforcement Learning: A Survey by Kaelbling, L.P., Littman, M.L., and Moore, A.W, 1996.
Additional reading:
- Reinforcement Learning: An Introduction by Richard Sutton
- Neuro-Dynamic Programming by Dimitri P. Bertsekas and John N. Tsitsiklis
- Reinforcement Learning: A Tutorial , M. Harmon and S. Harmon
- How to Make Software Agents Do the Right Thing: An Introduction to Reinforcement Learning, by Satinder Singh, Peter Norvig and David Cohn 1996.
Software & Links:
- Code examples from the book Reinforcement Learning in LISP, C, C++
- Reinforcement Learning software
- Reinforcement Learning Simulator by Tor Aamodt
Lecture 9: Evolutionary Algorithms : 29-11-2002 & 03-12-2002
Slides:
Homework:
- read Mitchell : chapter 9
- read : H.-P. Schwefel, T. Bäck, An Overview of Evolutionary Algorithms for Parameter Optimization, 1993.
Additional reading:
- T. Bäck, H.-P. Schwefel, Evolution Strategies I: Variants and their computational implementation, 1995.
- T. Bäck, H.-P. Schwefel, Evolution Strategies II: Theoretical aspects, 1995.
Software & Links:
Lecture 10: Learning Theory : 03-12-2002 & 05-12-2002
Slides:
Homework:
- read Mitchell : chapter 7
Additional reading:
- Computational Learning Theory in Algorithms and Theory of Computation Handbook, CRC Press, 1999. (Goldman).
Software & Links:
- COLT Computational Learning Theory
Lecture 11: Exam preparation : 10-12-2002
Slides:
Lecture 12: Fuzzy Logic & Robot Learning : 12-12-2002
Slides:
Homework:
- read : S. Schaal Robot Learning in Arbib, M. (ed.). Handbook of Brain Theory and Neural Networks, 2nd Edition. Cambridge, MA: MIT Press, (2002)
Additional reading:
- S. Schaal Is imitation learning the route to humanoid robots? Trends in Cognitive Sciences, 3(6), 233-242, (1999)
- C.G. Atkeson, A. Moore, S. Schaal Locally weighted learning for control" Artificial Intelligence Review, 11, 1-5, pp.75-113, (1997)
Software & Links:
- Movies robot learning

Software & Datasets & Links:

for further questions, contact Frank Hoffmann (phone: 790-6271)