MACHINE LEARNING

1
MACHINE LEARNING

• Fatemeh Saremi
• Mina Razaghpour

2
Overview

• What is Learning?
• What is Machine Learning?
• Why should Machines Learn?
• How machines learn?
• Specific Machine Learning Methods
• Solving Traveling Salesman Problem with Ant
  colony
• Summary
• References

3
What is Learning?

• Learning is any change in a system that allows it
  to perform better the second time on repetition
  of the same task or on another task drawn from
  the same population.
• One part of learning is acquiring knowledge and
  new information
• And the other part is problem-solving .

4
What is Machine Learning?

• The goal of machine learning is to build computer
  systems that can adapt and learn from their
  experience.
• Machine Learning algorithms discover the
  relationships between the variables of a system
  (input, output and hidden) from direct samples of
  the system.

System
. .
5
Why Should Machines Learn?

• We expect machines to learn from their mistakes.
• An Intelligence that didnt learn ,would not be
  much of an Intelligence.
• Machine Learning is a prerequisite for any mature
  programme of Artificial Intelligence.

6
How Machines Learn?

• Machine Learning typically follows three phases
• Training
• A training set of examples of correct behavior
  is analyzed and some representation of the newly
  learnt knowledge is stored. This is some form of
  rules.

7
How Machines Learn? (cont.)

• Validation
• The rules are checked and ,if necessary
  ,additional training
• is given . Sometimes additional test data are
  used , but instead , a human expert may validate
  the rules , or some other automatic
  knowledge - based component may be used. . The
  role of the tester is often called the critic.
• Application
• The rules are used in responding to some new
  situation.

8
How Machines Learn? (cont.)
Training set
Existing knowledge
Training
Test data
New knowledge
Validation
New situation
Critic
Application
Response
9
Specific Machine Learning Methods
10
Learning by Memorizing

• The simplest way of learning
• Storing examples of correct behavior
• An example
• Learn to play Checkers written by Samuel

11
Checkers

• Using min-max method.
• When complete search is impossible , use a Static
  Evaluation Function .
• At the end of each turn , Record computed values
  for each state.
• Reaching a state ,visited in previous games,
  stop the search and use the stored value.

12
Learning by Memorizing (cont.)

• It is too simple , and it is not sufficient for
  complicated problems.
• We also need
• Organized information storing
• Generalization
• Direction
• So in this method learning is similar to problem
  solving , but its success is dependent on proper
  structure for knowledge-base.

13
Learning by Adjusting Parameters

• Determining parameters
• Assigning initial weight to each parameter
• Modifying weights as the program goes on
• In Checkers
• 16 parameters for each state
• f c1t1 c2t2 c16t16
• When to modify a coefficient?
• And to what degree?

14
Learning by Adjusting Parameters(cont.)

• So in this method learning is similar to other
  problemsolving methods,and it is dependent
• on searching algorithms.

15
Learning by Exploration

• This program explores domains , looking
• for interesting patterns and generalizations.
• A remarkable program AM developed by
• Doug Lenat
• AM works in the domain of elementary mathematics.
• It maintains a large , growing database of
  concepts , such as set and function in the
  mathematics domain.

16
Learning by Exploration (cont.)

• The program maintains an agenda of tasks ,
• and keeps them stored in decreasing order
• of interestingness . Its basic control cycle
  is
• to select the first task from the agenda,work
• on it (which may add new tasks), and repeat.
• Working on a task is done by rules called
  heuristics .

17
Learning by Exploration (cont.)

• Another example is Eurisko ,developed by Doug
  Lenat
• Eurisko works in a variety of domains , including
  three-dimensional VLSI circuits and the design of
  battle fleets for a space warfare game.
• Eurisko is more complex than AM , and was
  designed to overcome some of AM s flaws. But
  both programs operate similarly.

18
Ant Colony SystemA Learning Approach to TSP
19
Ant Colony Algorithms

• Inspired from Ants Natural behavior
• Ants can find the shortest path between two
  points.
• However, they cant see! So How?

20
Finding the shortest path

• Ants choose paths according to amount of
  pheromone.
• Pheromone is accumulated faster on shorter path.
• After some time ,all of ants choose the shorter
  path.

21
Natural behavior of ant
22
ACS for Traveling Salesman Problem

• Having a set of simple agents called ants
• Each edge has a desirability measure called
  Pheromone
• Ants search in parallel for good solutions to TSP
• Ants Cooperate through pheromone-mediated
  communication

23
Algorithm

• Initialize randomly place ants in cities
• Each ant constructs a tour iteratively
• It chooses the next city by
• A Greedy Heuristic the nearest city
• Use past experience the Edge with Highest
  Pheromone Level

24
Updating Pheromone Level

• Global Updating At the end of each round
• The best solutions get extra point
• Local Updating

25
Algorithm

• Loop
• randomly place m artificial ants on n cities
• For city1 to n
• For ant1 to m
• select probabilistically the next
  city according
• to exploration and
  exploitation
• apply the local updating rule
• End For
• End For
• Apply the global updating rule using the best
  ant
• Until End_condition

26
Transition function
With Probability q0 Exploitation
With Probability (1- q0) Exploration
27
A simple TSP example
A
B
C
D
E
dAB 100dBC 60dDE 150
28
Iteration 1
A
B
C
D
E
29
How to build next sub-solution?
A
B
C
D
E
30
Iteration 2
A
B
C
D
E
31
Iteration 3
A
B
C
D
E
32
Iteration 4
A
B
C
D
E
33
Iteration 5
A
B
C
D
E
34
Path and Pheromone Evaluation
L1 300
L2 450
L3 260
L4 280
L5 420
35
Global Pheromone Updating

• Only the ant that generated the best tour is
  allowed to globally update the amount of
  pheromone on its tour edges.

36
Local Pheromone Updating

• If edge (r,s) is visited by ant

37
Effect of the Local Rule
Local update rule makes the edge pheromone level
diminish.
Visited edges are less less attractive as they
are visited by the various ants.
Favors exploration of not yet visited edges.
This helps in shuffling the cities so that
cities visited early in one ants tours are being
visited later in another ants tour.
38
Enhancements to ACS

• The Algorithm can be performed in Parallel, so
  the order is independent of ants number.
• For each size of problem a special set of values
  for ants number ant other parameters lead the
  best result.

39
Compare results with some well-known Algorithms
Problem name ACS GA EP SA optimum
Oliver30 (30-city) 420 830 421 3200 420 40000 424 24617 420
Eil50 (50-city 425 1830 428 25000 426 100000 443 68512 425
Eil75 (75-city 535 3480 545 80000 542 325000 580 173250 535
Kroa100 (100city) 21,282 4820 21,761 103000 N/A N/A 21,282
40
SUMMARY

• The goal of machine learning is to build computer
  systems
• that can adapt and learn from their
  experience.
• An Intelligence that didnt learn ,would not be
  much of an Intelligence.
• Machine Learning typically follows three phases
• Training
• Validation
• Application
• Specific Machine Learning Methods
• Learning by Memorizing
• Learning by Adjusting Parameters
• Learning by Exploration
• Ant colony algorithms
• an efficient ,nature-inspired learning algorithm
  for TSP

41
References

• J. Finlay A. Dix , An introduction to
  Artificial Inteligence , 1997.
• R.S. Michalaski J.G. Carbonell T.M.Mitchell ,
  Machine Learning ,1983.
• E. Charniak D. McDermott , Introduction to
  Artificial Inteligence , 1985.
• M. Fahimi , Artificial Inteligence , 2002.
• M.Dorigo,L.Gambadrella A Cooperative learning
  approach to the travelling salesman problem,IEEE
  Transactions,1997
• L.Gambadrella,M.Dorigo Ant Colonies for the
  travelling salesman problem,1997
• V. Maniezzo, L. Gambardella, F. de Luigi Ant
  colony Optimization,2001

42
QUESTIONS???

• Thanks for your attention!