Machine Learning 2D5362

Lecture 1:
Introduction to Machine Learning
Machine Learning
• Date/Time:
Tuesday ???
Thursday 13.30
• Location:
BB2 ?
• Course requirements:
active participation
homework assignments
course project
• Credits:
3-5 credits depending on course project
• Course webpage:
http://www.nada.kth.se/~hoffmann/ml.html
 Course Material
Textbook (recommended):
• Machine Learning
Tom M. Mitchell, McGraw Hill,1997
ISBN: 0-07-042807-7 (available as paperback)
Further readings:
• An Introduction to Genetic Algorithms
Melanie Mitchell, MIT Press, 1996
• Reinforcement Learning – An Introduction
Richard Sutton, MIT Press, 1998
• Selected publications:
check course webpage
 Course Overview
 Introduction to machine learning
 Concept learners
 Decision tree learning
 Neural networks
 Evolutionary algorithms
 Instance based learning
 Reinforcement learning
 Machine learning in robotics
 Software Packages & Datasets
• MLC++
• Machine learning library in C++
• http://www.sig.com/Technology/mlc

• GALIB
• MIT GALib in C++
• http://lancet.mit.edu/ga

• UCI
• Machine Learning Data Repository UC Irvine
• http://www.ics.uci.edu/~mlearn/ML/Repository.html
 Possible Course Projects
 Apply machine learning techniques to your own
problem e.g. classification, clustering, data
modeling, object recognition
 Investigating combining multiple classifiers
 Comparing different approaches in genetic fuzzy
systems
 Learning robotic behaviors using evolutionary
techniques or reinforcement learning
 LEGO Mindstorm
 Scout
 Scout Robots

 16 Sonar sensors
 Laser range
scanner
 Odometry
 Differential drive
 Simulator
 API in C
 LEGO Mindstorms

 Touch sensor
 Light sensor
 Rotation sensor
 Video cam
 Motors
 Learning & Adaptation

• ”Modification of a behavioral tendency by expertise.”

(Webster 1984)
• ”A learning machine, broadly defined is any device whose
actions are influenced by past experiences.” (Nilsson 1965)
• ”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.” (Simon 1983)
• ”An improvement in information processing ability that results
from information processing activity.” (Tanimoto 1990)
Learning

Definition:
A computer program is said to learn from
experience E with respect to some class
of tasks T and performance measure P,
if its performance at tasks in T, as
measured by P, improves with
experience.
Disciplines relevant to ML
 Artificial intelligence
 Bayesian methods
 Control theory
 Information theory
 Computational complexity theory
 Philosophy
 Psychology and neurobiology
 Statistics
 Applications of ML
 Learning to recognize spoken words
 SPHINX (Lee 1989)
 Learning to drive an autonomous vehicle
 ALVINN (Pomerleau 1989)
 Learning to classify celestial objects
 (Fayyad et al 1995)
 Learning to play world-class backgammon
 TD-GAMMON (Tesauro 1992)
 Designing the morphology and control structure
of electro-mechanical artefacts
 GOLEM (Lipton, Pollock 2000)
 Artificial Life
 GOLEM Project (Nature: Lipson, Pollack 2000)
http://golem03.cs-i.brandeis.edu/index.html
 Evolve simple electromechanical locomotion machines from
basic building blocks (bars, acuators, artificial neurons) in a
simulation of the physical world (gravity, friction).
 The individuals that demonstrate the best locomotion
ability are fabricated through rapid prototyping technology.
Evolvable Robot
Arrow
Ratchet
Tetra
 Evolved Creatures
Evolved creatures: Sims (1994)
http://genarts.com/karl/evolved-virtual-creatures.html
Darwinian evolution of virtual block creatures for
swimming, jumping, following, competing for a block
 Learning Problem

Learning: improving with experience at some task

 Improve over task T

 With respect to performance measure P

 Based on experience E

Example: Learn to play checkers:

 T: play checkers

 P: percentage of games won in a tournament

 E: opportunity to play against itself

Learning to play checkers

 T: play checkers
 P: percentage of games won
 What experience?
 What exactly should be learned?
 How shall it be represented?
 What specific algorithm to learn it?
 Type of Training Experience
 Direct or indirect?
 Direct: board state -> correct move
 Indirect: outcome of a complete game
 Credit assignment problem
 Teacher or not ?
 Teacher selects board states
 Learner can select board states
 Is training experience representative of
performance goal?
 Training playing against itself
 Performance evaluated playing against world champion
 Choose Target Function
 ChooseMove : B  M : board state  move
 Maps a legal board state to a legal move
 Evaluate : BV : board state  board value
 Assigns a numerical score to any given board
state, such that better board states obtain a
higher score
 Select the best move by evaluating all successor
states of legal moves and pick the one with the
maximal score
Possible Definition of Target
Function
 If b is a final board state that is won then
V(b) = 100
 If b is a final board state that is lost then
V(b) = -100
 If b is a final board state that is drawn then
V(b)=0
 If b is not a final board state, then
V(b)=V(b’), where b’ is the best final board
state that can be achieved starting from b
and playing optimally until the end of the
game.
 Gives correct values but is not operational
 State Space Search
V(b)= ?

V(b)= maxi V(bi)

m1 : bb1 m2 : bb2 m3 : bb3

 State Space Search
V(b1)= ?

V(b1)= mini V(bi)

m4 : bb4 m5 : bb5 m6 : bb6

Final Board States

Black wins: V(b)=-100

Red wins: V(b)=100

draw: V(b)=0
Depth-First Search
Breadth-First Search
 Number of Board States
Tic-Tac-Toe:
#board states < 9!/(5! 4!) + 9!/(1! 4! 4!) + …
… + 9!/(2! 4! 3!) + … 9 = 6045

4 x 4 checkers: (no queens)

#board states = ?

#board states < 8x7x6x5*22/(2!*2!) = 1680

Regular checkers (8x8 board, 8 pieces each)

#board states < 32!*216/(8! * 8! * 16!) = 5.07*1017
Choose Representation of
Target Function
 Table look-up
 Collection of rules
 Neural networks
 Polynomial function of board features
 Trade-off in choosing an expressive
representation:
 Approximation accuracy
 Number of training examples to learn the
target function
 Representation of Target
Function
V(b)=0 + 1bp(b) + 2rp(b) +
3bk(b) + 4rk(b) + 5bt(b) + 6rt(b)
 bp(b): #black pieces

 rb(b): #red pieces

 bk(b): #black kings

 rk(b): #red kings

 bt(b): #red pieces threatened by black

 rt(b): #black pieces threatened by red

Obtaining Training Examples

 V(b) : true target function

 V’(b) : learned target function
 Vtrain(b) : training value
 Rule for estimating training values:
 Vtrain(b)  V’(Successor(b))
Choose Weight Training Rule
LMS weight update rule:
 Select a training example b at random
1. Compute error(b)
error(b) = Vtrain(b) – V’(b)
2. For each board feature fi, update
weight i  i +  fi error(b)
 : learning rate approx. 0.1
 Example: 4x4 checkers
V(b)=0 + 1rp(b) + 2bp(b)
Initial weights: 0=-10, 1 =75, 2 =-60

V(b0)=0 + 1*2 + 2*2 = 20

m1 : bb1 m2 : bb2 m3 : bb3

V(b1)=20 V(b2)=20 V(b3)=20
 Example 4x4 checkers

V(b0)=20 V(b1)=20

1. Compute error(b0) = Vtrain(b) – V(b0) = V(b1) – V(b0) = 0

2. For each board feature fi, update weight
i  i +  fi error(b)
0  0 + 0.1 * 1 * 0
1  1 + 0.1 * 2 * 0
2  2 + 0.1 * 2 * 0
Example: 4x4 checkers

V(b0)=20 V(b1)=20

V(b3)=20 V(b2)=20
Example: 4x4 checkers

V(b3)=20

V(b4a)=20 V(b4b)=-55
 Example 4x4 checkers

V(b3)=20 V(b4)=-55

1. Compute error(b3) = Vtrain(b) – V(b3) = V(b4) – V(b3) = -75

2. For each board feature fi, update weight

i  i +  fi error(b) : 0=-10, 1 =75, 2 =-60

0  0 - 0.1 * 1 * 75, 0 = -17.5
1  1 - 0.1 * 2 * 75, 1 = 60
2  2 - 0.1 * 2 * 75, 2 = -75
Example: 4x4 checkers
0 = -17.5 , 1 = 60, 2 = -75

V(b4)=-107.5 V(b5)=-107.5
 Example 4x4 checkers

V(b5)=-107.5 V(b6)=-167.5

error(b5) = Vtrain(b) – V(b5) = V(b6) – V(b5) = -60

0=-17.5, 1 =60, 2 =-75
i  i +  fi error(b)
0  0 - 0.1 * 1 * 60, 0 = -23.5
1  1 - 0.1 * 1 * 60, 1 = 54
2  2 - 0.1 * 2 * 60, 2 = -87
 Example 4x4 checkers
Final board state: black won Vf(b)=-100

V(b6)=-197.5

error(b6) = Vtrain(b) – V(b6) = Vf(b6) – V(b6) = 97.5

0=-23.5, 1 =54, 2 =-87
i  i +  fi error(b)
0  0 + 0.1 * 1 * 97.5, 0 = –13.75
1  1 + 0.1 * 0 * 97.5, 1 = 54
2  2 + 0.1 * 2 * 97.5, 2 = -67.5
 Evolution of Value Function

Training data:
before
after
 Design Choices
Determine Type of
Training Experience
Games against Games Table of correct
experts against self moves
Determine
Target Function
BoardMove BoardValue
Determine Representation
of Learned Function
polynomial Linear function of Artificial neural
six features network
Determine
Learning Algorithm
Gradient descent Linear programming
 Learning Problem Examples
 Credit card applications
 Task T: Distinguish ”good” applicants from
”risky” applicants.
 Performance measure P : ?
 Experience E : ? (direct/indirect)
 Target function : ?
 Performance Measure P:
 Error based: minimize percentage of incorrectly
classified customers : P = Nfp + Nfn / N
Nfp: # false positives (rejected good customers)
Nfn: # false negatives (accepted bad customers)

 Utility based: maximize expected profit of credit

card business: P = Ncp *Ucp+ Nfn *Ufn
Ucp : expected utility of an accepted good customer
Ufn : expected utility/loss of an accepted bad customer
 Experience E:
 Direct: Decisions on credit card applications
made by a human financial expert
Training data: <customer inf., reject/accept>
 Direct: Actual customer behavior based on
previously accepted customers
Training data: <customer inf., good/bad>
Problem: Distribution of applicants Papplicant is not
identical with training data Ptrain
 Indirect: Evaluate a decision policy based on the
profit you made over the past N years.
 Distribution of Applicants

Good customers Cw=38

Bad customers
Assume we want to minimize
classification error:
What is the optimal decision
boundary?
 Distribution of Accepted
Customers
Cw=43
Good customers
Bad customers
What is the optimal decision
boundary?
 Target Function

Customer record:
income, owns house, credit history, age, employed, accept
$40000, yes, good, 38, full-time, yes
$25000, no, excellent, 25, part-time, no
$50000, no, poor, 55, unemployed, no
 T: Customer data  accept/reject
 T: Customer data  probability good customer
 T: Customer data  expected utility/profit
 Learning methods
 Decision rules:
 If income < $30.000 then reject
 Bayesian network:
 P(good | income, credit history,….)
 Neural Network:
 Nearest Neighbor:
 Take the same decision as for the customer
in the data base that is most similar to the
applicant
 Learning Problem Examples
 Obstacle Avoidance Behavior of a
Mobile Robot
 Task T: Navigate robot safely through an
environment.
 Performance measure P : ?
 Experience E : ?
 Target function : ?
 Performance Measure P:

 P: Maximize time until collision with

obstacle
 P: Maximize distance travelled until
collision with obstacle
 P: Minimize rotational velocity, maximize
translational velocity
 P: Minimize error between control
action of a human operator and robot
controller in the same situation
 Training Experience E:
 Direct: Monitor human operator and use
her control actions as training data:
 E = { <perceptioni, actioni>}
 Indirect: Operate robot in the real world
or in a simulation. Reward desirable
states, penalize undesirable states
 V(b) = +1 if v > 0.5 m/s
 V(b) = +2 if  < 10 deg/s
 V(b) = -100 if bumper state = 1

Question: Internal or external reward ?

 Target Function
 Choose action:
 A: perception  action

Sonar readings: s1(t)…sn(t)  <v,>

 Evaluate perception/state:
 V: s1(t)…sn(t)  V(s1(t)…sn(t))

 Problem: states are only partially observable

therefore world seems non-deterministic
 Markov Decision Process : successor state s(t+1) is
a probabilistic function of current state s(t) and
action a(t)
 Evaluate state/action pairs:
 V: s1(t)…sn(t), a(t)  V(s1(t)…sn(t),a(t))
Learning Methods

 Neural Networks
 Require direct training experience
 Reinforcement Learning
 Indirect training experience
 Evolutionary Algorithms
 Indirect training experience
 Evolutionary Algorithms
mutation population of genotypes

10111
10011 10001
phenotype space
00111
01001
01001 11001
01011 f

coding scheme
recombination selection
x
10011
10
10001
011
001 10001
fitness
01001
01
01011
001
011 10001 11001
01011
Evolution of Simple Navigation
 Issues in Machine Learning
 What algorithms can approximate functions
well and when?
 How does the number of training examples
influence accuracy?
 How does the complexity of hypothesis
representation impact it?
 How does noisy data influence accuracy?
 What are the theoretical limits of learnability?
 Machine vs. Robot Learning
Machine Learning Robot Learning
 Learning in vaccum  Embedded learning
 Statistically well-behaved  Data distribution not
data homegeneous
 Mostly off-line  Mostly on-line
 Informative feed-back
 Qualitative and sparse
 Computational time not an feed-back
issue
 Time is crucial
 Hardware does not matter
 Convergence proof  Hardware is a priority
 Empirical proof
 Learning in Robotics
• behavioral adaptation:
adjust the parameters of individual behaviors according
to some direct feedback signal (e.g. adaptive control)
• evolutionary adaptation:
application of artificial evolution to robotic systems
• sensor adaptation:
adopt the perceptual system to the environment
(e.g. classification of different contexts, recognition)
• learning complex, deliberative behaviors:
unsupervised learning based on sparse feedback from
the environment, credit assignment problem
(e.g. reinforcement learning)