Transcript Machine Learning

Machine Learning
System Modelling and Analysis
Burak Tiftik
1
Summary
●
What is Machine Learning
●
History of Machine Learning
●
Problem Types in Machine Learning
●
Types of Algorithms Used
●
Example of Machine Learning in Action
●
Conclusion
●
Questions
2
What is Machine Learning
●
Programs that can learn from data without
explicit programming.
Example:
●
Car driving itself
●
Program that estimates house prices
●
Spam filter
3
Daily Examples
●
●
Mailing office automatic zipcode reader
Bank applications on understanding
unathorized usage of cards
●
Online shopping suggestion methods
●
Fingerprint recognition
●
Face recognition
●
Handwriting recognition
4
History of Machine Learning
●
●
●
●
Checkers 1952 Arthur
Samuel, IBM for IBM
701
Beats Connecticut
champion 1962
Who then beats the
program 6 times in a
row!
Alpha vs. Beta
approach
5
History of Machine Learning
●
1957 Perceptron
by Frank Rosenblatt
in Cornell Aeronautical
Lab
●
●
A type of neural
network
1969 Martin Minsky
publishes a paper
finding faults in
perceptrons
6
History of Machine Learning
●
●
1967 Pattern
Recognition
First algorithm is
K-NN!
7
History of Machine Learning
●
1970's were not great for Machine Learning
●
Time of Expert Systems
8
History of Machine Learning
●
●
1980 Decision Tree
model becomes
available
Understandable by
humans
9
History of Machine Learning
●
●
●
1980 Multi-Layered
Neural Networks
Perceptrons
upgraded
Can solve much
complex problems
10
History of Machine Learning
●
●
●
1995 State Vector
Machine (SVM)
Vladimir N. Vapnik
and Corrina Cortes
Used in text
recognition,
classification of
messages
Classification and
regression
11
Problems, Algorithms Types, Algorithms
●
There is a disctinction between terminology
●
Problem?
●
Algorithm Type?
●
Algorithm?
12
Types of Problems
●
Classification
●
Regression
●
Clustering
●
And many more
13
Classification
●
●
●
●
Distinguishing
between different
things
Provided we know
what these things are
Example: Character
recognition
Methods: Support Vector
Machines, Neural
Networks, Naive Bayes
classifier, Decision
Trees, K-NN
14
Regression
●
●
●
Predicting future input
results from previous
results
Example: House
prices
Methods: Linear
Regression, Non-Linear
Regression, Generalized
Linear Models, Decision
Trees, Neural Networks
15
Clustering
●
●
●
Given random data,
grouping things by
similarity
Example: Sort books
in library by content
Methods: Hierarchical
Clustering, k-means
Clustering, Gaussian
Mixture Models, SelfOrganizing Maps
16
Types of Algorithms
●
Supervised Learning
●
Unsupervised Learning
●
Transduction
●
Reinforcement Learning
●
And many more
17
Supervised Learning
●
A training data is available
●
Training data contains input and desired output
●
Once trained, it is ready to use
●
Used in Classification and Regression
●
Methods: Decision Trees, K-NN, Linear Regression,
Naive Bayes classifier, Neural Networks, Support
Vector Machines, Case Based Reasoning ...
18
Unsupervised Learning
●
Training data is not available or not used
●
Focus is on seperation of raw data into parts
●
Used in Clustering
●
Methods: K-means, hidden Markov models, SelfOrganizing Maps, Hierarchical Clustering ...
19
Example
●
●
●
●
●
Learning to win: case based plan selection in
real-time strategy game
By David W. Aha, Matthew Molineaux and Marc
Ponsen
Implemented in Wargus, clone of Warcraft II
Uses Case Based Reasoning, which is a
method in supervised learning
Beats all available script based AI
20
Wargus
●
Real Time Strategy
game
●
Open Source
●
2 sides fight another
●
Buildings, units
21
How It Works
●
●
●
Buildings are used to
calculate possible
production
Ability to produce
different units dictate
AI move
Learns in time
22
Quick Refresher
Case: Log 11
Case: Current Case
Computer Heat: 120
Computer Heat: 130
Computer Speed: 60
Computer Speed: 55
Does computer turn on?: Yes
Does computer turn on?: Yes
Does sound work?: Yes
Does sound work?: Yes
Hard Disk Remaining: 40 gb
Hard Disk Remaining: 6 gb
●
●
●
Red Text implies distance between two cases.
Black text is weighted more than any other as
problem itself is speed.
Sound is weighted little for instance as it is
unrelated
23
How It Works
●
Variables in CBR Algorithm
24
Euclidean Distance Refresher
Calculates similarity between points using distance
25
Best Case Selector
●
C is past case
●
S is new case
●
Priority is on distance and secondarily on
performance (similarity)
26
What Do These Mean?
27
What Do These Mean 2?
28
Conclusion
●
●
●
●
Machine Learning is a huge field that is
concerned with learning from data
It has huge impact on our lives even without our
knowing
Has a lot of usage in business world where
large amounts of data is common place
Exciting field!
29
Questions?
30