1. Machine Learning
CSE 4308/5360 – Artificial Intelligence I
Darin Brezeale
The University of Texas at Arlington
Machine Learning – p. 1/2
2. Introduction
Machine learning is the area of artificial intelligence
in which we ‘teach’ the computer what it knows. This
is typically done by providing training examples,
either labeled (supervised learning) or unlabeled
(unsupervised learning).
Many different methods have been developed for
performing machine learning, each with its strengths
and weaknesses.
Machine Learning – p. 2/2
3. Introduction
Why learn instead of just hard-coding rules?
• It may not be practical; the number of rules may
be too large.
• Humans may not understand the domain
well-enough to write the rules, e.g., in
recognizing faces.
• Things change; we would like a program to be
able to change as well without much human
intervention.
Machine Learning – p. 3/2
4. Data Representation
A question that is often neglected when discussing
machine learning is ‘How do I represent my training
examples?’
A typical approach is to represent the examples as
vectors.
EX: If we had samples for various animals, then the
sample
<dog, has hair, barks, flys, tail, no gills>
might become
{dog, 1, 1, 0, 1, 0}
Machine Learning – p. 4/2
5. Neural Network
A neural network (or artificial neural network) is a
network of simple units.
The artificial neural network is inspired by the
network of neurons in the human brain.
Neural networks can handle discrete or real-valued
functions and have been very popular in various
pattern recognition applications, e.g., recognizing
faces.
Machine Learning – p. 5/2
7. Neural Network
The value of a unit is a weighted sum of its input.
Figure 2: Relationship between a single unit and its input units.
Machine Learning – p. 7/2
8. Neural Network
The big questions when using a neural network are:
1. What type of network (e.g., feedforward,
back-propagation, etc.) should be used?
2. What should the network topology (i.e., how
many hidden nodes) be?
Machine Learning – p. 8/2
9. Neural Network
Pros:
• can fit nonlinear functions
• can handle noisy data
Cons:
• hard to understand the resulting function weights
• can take a long time to train
Machine Learning – p. 9/2
10. Genetic Algorithm
Genetic algorithms are based on the concept of
natural selection with the attributes of the samples
treated as ‘chromosomes’.
They are useful for solving optimization problems.
We don’t train the system; instead, we are seeking the
combination of values that produces the best result.
Machine Learning – p. 10/2
11. Genetic Algorithm
The various solutions to a problem are individuals in a
population.
Some produce better function values than others.
The stronger individuals are combined to produce new
offspring with operations such as crossover and
mutation.
Machine Learning – p. 11/2
12. Genetic Algorithm
Figure 3: Application of crossover.
Machine Learning – p. 12/2
13. Genetic Algorithm
Pros:
• Domain independent.
• Can overcome local optima.
Cons:
• Not always clear how to represent the problem.
• Must choose fitness function, rate of cross-over,
rate of mutation, etc.
Machine Learning – p. 13/2
14. Decision Tree
Decision trees are trees in which the branches
represent the values of attributes.
They can be used to represent discrete-valued
functions.
Machine Learning – p. 14/2
15. Decision Tree
Figure 4: Simplistic decision tree for classifying some types of
animals.
Machine Learning – p. 15/2
16. Decision Tree
The big question when constructing decision trees is
the order of attributes to branch on. Often the
attributes are ordered by the information gain from
them.
Example: If we constructed a decision tree to classify
mammals as dog or not-dog, the top split should
not be ‘walks’ since most mammals walk. A better
choice might be ‘barks’.
Machine Learning – p. 16/2
17. Decision Tree
Pros: easy to understand results
Cons: prone to overfitting (pruning can address this)
Machine Learning – p. 17/2
18. Reinforcement Learning
In reinforcement learning, the agent receives feedback
from its actions and seeks to maximize its total
reward.
This is different from supervised learning in that
training examples are not provided.
Used for sequential decisions.
See examples at:
http://www.cs.ualberta.ca/ sutton/book/ebook/node8.html
Machine Learning – p. 18/2
19. More Recent Methods
The methods previously described are some of the
oldest. More recent methods are:
• statistics-based – actually, these have been around
for a while, but they remain the focus of much
research, especially as statisticians begin looking
at machine learning and data mining
• matrix decompositions – learn a basis for each
category; classify as category whose basis vectors
best represent the new sample.
Machine Learning – p. 19/2
20. References
• Hastie, Trevor, Robert Tibshirani, and Jerome Friedman, The Elements
of Statistical Learning, 2nd ed, Springer, 2009.
• Marczyk, Adam, Genetic Algorithms and Evolutionary Computation,
URL: http://www.talkorigins.org/faqs/genalg/genalg.html, accessed:
August 9, 2009.
• Mitchell, Tom M., Machine Learning, McGraw-Hill, 1997.
• Moore, Andrew W., Statistical Data Mining Tutorials, URL:
http://www.autonlab.org/tutorials/, accessed: August 9, 2009.
• Sutton, Richard S. and Andrew G. Barto, Reinforcement Learning: An
Introduction, The MIT Press, 1998.
Machine Learning – p. 20/2