Concept

Learning
Waiting outside the house to get an autograph.

2
Which days does he come out to enjoy sports?
• Sky condition
• Humidity
• Temperature
• Wind
• Water
• Forecast

• Attributes of a day: takes on values

3
Learning Task
• We want to make a hypothesis about the day on which
SRK comes out..
– in the form of a boolean function on the attributes of the
day.

• Find the right hypothesis/function from historical data

4
 Training Examples for EnjoySport

c( )=1
c( )=1
c( )=0
c( )=1

 Negative and positive learning

examples
 Concept c is the target
- Deriving a Boolean function from
learning: concept
training
examples
- Many “hypothetical” boolean functions
 Hypotheses; find h such that h = c.

– Other more complex examples:

 Non-boolean functions

 Generate hypotheses for concept from TE’s 5

 Representing Hypotheses
• Task of finding appropriate set of hypotheses for concept given training data
• Represent hypothesis as Conjunction of constraints of the following form:
– Values possible in any hypothesis
 Specific value : Water  Warm
 Don’t-care value: Water  ?
 No value allowed : Water  
– i.e., no permissible value given values of other attributes
– Use vector of such values as hypothesis:
  Sky AirTemp Humid Wind Water Forecast 
– Example: Sunny ? ? Strong ?
Same 
• Idea of satisfaction of hypothesis by some example
– say “example satisfies hypothesis”
– defined by a function h(x):
h(x)   if h is true on x
• Want hypothesisthat
 otherwise
best fits examples: 6
– Can reduce learning to search problem over space of
 Prototypical Concept Learning Task
TASK T: predicting when person will enjoy sport
– Target function c: EnjoySport : X   
– Cannot, in general, know Target function c
 Adopt hypotheses H about c
– Form of hypotheses H:
 Conjunctions of literals ?, Cold, High, ?, ?, ? 
 EXPERIENCE E
– Instances X: possible days described by attributes Sky, AirTemp,
Humidity,
Wind, Water, Forecast
– Training examples D: Positive/negative examples of target function {x1,
cx1, . . . xm, cxm}
 PERFORMANCE MEASURE P: Hypotheses h in H such that hx = cx for
all x
in D ()
– There may exist several alternative hypotheses that fit examples
7
 Inductive Learning Hypothesis

Any hypothesis found to approximate the target

function well over a sufficiently large set of
training examples will also approximate the target
function well over other unobserved examples

8
 Approaches to learning algorithms
The choice of the
• Brute force search hypothesis space reduces
– Enumerate all possible hypotheses and evaluate the number of
– Highly inefficient even for small EnjoySport examplehypotheses.
 |X| = 3.2.2.2.2= 96 distinct instances
 Large number of syntactically distinct hypotheses (0’s, ?’s)
– EnjoySport: |H| = 5.4.4.4.4.4=5120
– Fewer when consider h’s with 0’s
Every h with a 0 is empty set of instances (classifies instance as neg)
Hence # semantically distinct h’s is:
1+ (4.3.3.3.3.3) = 973
 EnjoySport is VERY small problem compared to many
• Hence use other search procedures.
– Approach 1: Search based on ordering of hypotheses
– Approach 2: Search based on finding all possible hypotheses using
a good representation of hypothesis space
 All hypotheses that fit data
9
 Ordering on Hypotheses
Instances X
Hypotheses H specific

general

xSunny Warm High Strong Cool hSunny ? ? Strong ? ?

Same xSunny Warm High Light Warm  hSunny ? ? ? ? ?
Same hSunny ? ? ? Cool ?
 h is more general than h( h g h) if for each
instance x, hx    hx  
 Which is the most general/most specific
 Find-S Algorithm
Assumes
There is hypothesis h in H describing target function c
There are no errors in the TEs
Procedure
1. Initialize h to the most specific hypothesis in H (what is this?)
2. For each positive training instance x
For each attribute constraint ai in h
If the constraint ai in h is satisfied by x
do nothing
Else
replace ai in h by the next more general constraint that is satisfied by x
3. Output hypothesis h
Note
There is no change for a negative example, so they are ignored.
This follows from assumptions that there is h in H describing target function c (ie.,for this h,
h=c) and that there are no errors in data. In particular, it follows that the hypothesis at any stage
cannot
be changed by neg example.

Assumption: Everything except the 11

positive examples is negative
 Example of Find-S
Instances Hypotheses H
X
specific

general

h     
xSunny Warm Normal Strong Warm Same hSunny Warm Normal Strong Warm
 xSunny Warm High Strong Warm Same  Same
xRainy Cold High Strong Warm Change  hSunny Warm ? Strong Warm
xSunny Warm High Strong Cool Change  Same hSunny Warm ? Strong
Warm Same hSunny Warm ?
Strong ? ?
Problems with Find-S
• Problems:
– Throws away information!
 Negative examples
– Can’t tell whether it has learned the concept
 Depending on H, there might be several h’s that fit TEs!
 Picks a maximally specific h (why?)
– Can’t tell when training data is inconsistent
 Since ignores negative TEs
• But
– It is simple
– Outcome is independent of order of examples
 Why?
• What alternative overcomes these problems?
– Keep all consistent hypotheses!
 Candidate elimination algorithm
13
Consistent Hypotheses and Version
Space
• A hypothesis h is consistent with a set of training examples
D of target concept c
if hx  cx for each training example x cx in D
– Note that consistency is with respect to specific D.
• Notation:
Consistent h, D  x cxD :: hx  cx
• The version space, VSH,D , with respect to hypothesis space
H and training examples D, is the subset of hypotheses
from H consistent with D
• Notation:
VSH,D = h | h  H  Consistent h, D

14
List-Then-Eliminate Algorithm
1. VersionSpace  list of all hypotheses in H
2. For each training example x cx
remove from VersionSpace any hypothesis h for which
hx  cx
3. Output the list of hypotheses in VersionSpace
4. This is essentially a brute force procedure

15
 Example of Find-S,
Revisited
Instances Hypotheses H
X
specific

general

h     
xSunny Warm Normal Strong Warm Same
hSunny Warm Normal Strong Warm

Same
Sunny Cold
xRainy WarmHigh
HighStrong
StrongWarm Same 
WarmChange
hSunny Warm ? Strong Warm
xSunny Warm High Strong Cool Change h Sunny
Same Warm ?Warm
hSunny Strong ? ?
? Strong

Warm Same
 Version Space for this Example
S
Sunny Warm ? Strong ? ? 

Sunny ? ? Strong ? ? Sunny Warm ? ? ? ? ? Warm ? Strong ? ?

G Sunny ? ? ? ? ?  ? Warm ? ? ? ?

17
Representing Version Spaces
• Want more compact representation of VS
– Store most/least general boundaries of space
– Generate all intermediate h’s in VS
– Idea that any h in VS must be consistent with all TE’s
 Generalize from most specific boundaries
 Specialize from most general boundaries

• The general boundary, G, of version space VSH,D is the set

of its maximally general members consistent with D
– Summarizes the negative examples; anything more general will
cover a negative TE
• The specific boundary, S, of version space VSH,D is the set
of its maximally specific members consistent with D
– Summarizes the positive examples; anything more specific will
fail
to cover a positive TE
18
 Theorem
Every member of the version space lies
between the S,G boundary
VSH,D  h | h  H  sS gG g  h  s
• Must prove:
– 1) every h satisfying RHS is in VSH,D;
– 2) every member of VSH,D satisfies RHS.
• For 1), let g, h, s be arbitrary members of G, H, S respectively
with g>h>s
– s must be satisfied by all + TEs and so must h because it is more general;
– g cannot be satisfied by any – TEs, and so nor can h
– h is in VSH,D since satisfied by all + TEs and no – TEs
• For 2),
– Since h satisfies all + TEs and no – TEs, h  s, and g  h. 19
Candidate Elimination Algorithm
G  maximally general hypotheses in H
S  maximally specific hypotheses in H
For each training example d, do
• If d is positive
– Remove from G every hypothesis
inconsistent with d
– For each hypothesis s in S that is
inconsistent with d
 Remove s from S
 Add to S all minimal
generalizations h of s such that
1. h is consistent with d, and
2. some member of G is more 20
general than h
 Candidate Elimination Algorithm
•
(cont)
If d is a negative example
– Remove from S every hypothesis inconsistent with d
– For each hypothesis g in G that is inconsistent with d
 Remove g from G
 Add to G all minimal specializations h of g such that
1. h is consistent with d, and
2. some member of S is more specific than h
– Remove from G every hypothesis that is less general than another
hypothesis in G
• Essentially use
– Pos TEs to generalize S
– Neg TEs to specialize G
• Independent of order of TEs
• Convergence guaranteed if:
– no errors
21
– there is h in H describing
 Example

Recall : If d is positive
Remove from G every hypothesis inconsistent
S0       with d
For each hypothesis s in S that is inconsistent
with d
• Remove s from S
G0 ? ? ? ? ? ? • Add to S all minimal generalizations h of s
that
are specializations of a hypothesis in G
• Remove from S every hypothesis
that
Sunny Warm Normal Strong Warm is more
Same  general than another
hypothesis in S

S1 Sunny Warm Normal Strong Warm Same

G1 ? ? ? ? ? ?
 Example (contd)

S1 Sunny Warm Normal Strong Warm Same

G1 ? ? ? ? ? ?

Sunny Warm High Strong Warm Same +

S2 Sunny Warm ? Strong Warm Same

G2 ? ? ? ? ? ?

23
 Example (contd)

S2 Sunny Warm ? Strong Warm Same Recall: If d is a negative

example
G2 ? ? ? ? ? ? – Remove from S every hypothesis inconsistent with d
– For each hypothesis g in G that is inconsistent with d
 Remove g from G
 Add to G all minimal specializations h of g that
generalize some hypothesis in S
 Remove from G every hypothesis that is less
another than
general hypothesis
in G
Rainy Cold High Strong Warm Change  Current G boundary is
S3 Sunny Warm ? Strong Warm Same incorrect So, need to make it
more specific.

G3
Sunny ? ? ? ? ? ? Warm ? ? ? ? ? ? ? ? ? Same 
24
 Example (contd)

 Why are there no hypotheses left relating

to:
–  Cloudy ? ? ? ? ?
 The following specialization using the third
value
? ? Normal ? ? ?,
Sunny Warm ? Strong Warm Same
is not more general than the specific
boundary
 The specializations ? ? ? Weak ? ?,
? ? ? ? Cool ? are also inconsistent
with S

25
 Example
(contd)
S3 Sunny Warm ? Strong Warm Same

G3 Sunny ? ? ? ? ? ? Warm ? ? ? ? ? ? ? ? ?Same

Sunny Warm High Strong Cool Change 

S4 Sunny Warm ? Strong ? ?

G4 Sunny ? ? ? ? ? ? Warm ? ? ? ?

26
 Example
(contd)Sunny Warm High Strong Cool Change 
 Why does this example remove a hypothesis from G?:

– ? ? ? ? ? Same

 This hypothesis
– Cannot be specialized, since would not cover new TE
– Cannot be generalized, because more general would cover
negative TE.
– Hence must drop hypothesis.

27
 Version Space of the Example
S Sunny Warm ? Strong ? ?

Sunny ? ? Strong ? ? Sunny Warm ? ? ? ? ? Warm ? Strong ? ?

S
G Sunny ? ? ? ? ? ? Warm ? ? ? ?
versio
n
space

G
28
 Convergence of algorithm
• Convergence guaranteed if:
– no errors
– there is h in H describing c.
• Ambiguity removed from VS when S = G
– Containing single h
– When have seen enough TEs
• If have false negative TE, algorithm will remove every h
consistent with TE, and hence will remove correct target concept
from VS
– If observe enough TEs will find that S, G boundaries converge to empty VS

29
 Let us try this

Origin Manufacturer Color Decade Type

Japan Honda Blue 1980 Economy +

Japan Toyota Green 1970 Sports -
Japan Toyota Blue 1990 Economy +

USA Chrysler Red 1980 Economy -

Japan Honda White 1980 Economy +

30
 And this

Origin Manufacturer Color Decade Type

Japan Honda Blue 1980 Economy +

Japan Toyota Green 1970 Sports -
Japan Toyota Blue 1990 Economy +
USA Chrysler Red 1980 Economy -
Japan Honda White 1980 Economy +
Japan Toyota Green 1980 Economy +
Japan Honda Red 1990 Economy -

31
 Which Next Training Example?
S Sunny Warm ? Strong ? ?

Sunny ? ? Strong ? ? Sunny Warm ? ? ? ? ? Warm ? Strong ? ?

G Sunny ? ? ? ? ? ? Warm ? ? ? ?

Order of
ssume learner can choose the next TE examples
• Should choose d such that matters for
– Reduces maximally the number of hypotheses in intermediate
sizes of S,G;
VS
not for the final
– Best TE: satisfies precisely 50% hypotheses;
S, G
 Can’t always be done
– Example:
 Sunny Warm Normal Weak Warm Same ? 32
 If pos, generalizes S
 Classifying new cases using VS

S Sunny Warm ? Strong ? ?

Sunny ? ? Strong ? ? Sunny Warm ? ? ? ? ? Warm ? Strong ? ?

G Sunny ? ? ? ? ? ? Warm ? ? ? ?

• Use voting procedure on following examples:

 Sunny Warm Normal Strong Cool Change
 Rainy Cool Normal Weak Warm Same
 Sunny Warm Normal Weak Warm Same
 Sunny Cold Normal Strong Warm Same
33
Effect of incomplete hypothesis
space
• Preceding algorithms work if target function is in H
– Will generally not work if target function not in H
• Consider following examples which represent target
function
“sky = sunny or sky = cloudy”:
 Sunny Warm Normal Strong Cool Change Y
 Cloudy Warm Normal Strong Cool Change Y
 Rainy Warm Normal Strong Cool Change N
• If apply CE algorithm as before, end up with empty
VS
– After first two TEs, S= ? Warm Normal Strong Cool
 it covers the third negative TE!
Change Need
– New hypothesis is overly general
• Our more
H does not include the appropriate expressiv
c e 34
hypothes
Incomplete hypothesis space
• If c not in H, then consider generalizing representation of H
to contain c
– For example, add disjunctions or negations to
representation of hypotheses in H
• One way to avoid problem is to allow all
possible
representations of h’s
– Equivalent to allowing all possible subsets of instances as
defining the concept of EnjoySport
 Recall that there are 96 instances in EnjoySport; hence there are
296 possible hypotheses in full space H
 Can do this by using full propositional calculus with AND, OR,
NOT
 Hence H defined only by conjunctions of attributes is biased
(containing only 973 h’s)
35
Unbiased Learners and Inductive
Bias
• BUT if have no limits on representation of hypotheses
(i.e., full logical representation: and, or, not), can only learn
examples…no generalization possible!
– Say have 5 TEs {x1, x2, x3, x4, x5}, with x4, x5 negative TEs
• Apply CE algorithm
– S will be disjunction of positive examples (S={x1 OR x2 OR
x3})
– G will be negation of disjunction of negative examples
(G={not (x4 or x5)})
– Need to use all instances to learn the concept!
• Cannot predict usefully:
– TEs have unanimous vote
– other h’s have 50/50 vote!
 For every h in H that predicts +, there is another that predicts -

36
Unbiased Learners and Inductive
Bias
• Approach:
– Place constraints on representation of
hypotheses
 Example of limiting connectives to conjunctions
 Allows learning of generalized hypotheses
 Introduces bias that depends on hypothesis representation
• Need formal definition of inductive bias of learning
algorithm

37
Inductive Syst and Equiv Deductive
Syst
• Inductive bias made explicit in equivalent deductive
system
– Logically represented system that produces same outputs
(classification) from inputs (TEs, instance x, bias B) as
CE procedure
• Inductive bias (IB) of learning algorithm L is any
minimal set of assertions B such that for any target
concept c and training examples D, we can logically infer
value c(x) of any instance x from B, D, and x
– E.g., for rote learner, B = {}, and there is no IB
• Difficult to apply in many cases, but a useful guide

38
Inductive Bias and specific learning
algs
• Rote learners:
no IB
• Version space candidate elimination algorithm:
c can be represented in H
• Find-S: c can be represented in H;
all instances that are not positive are
negative

39
Computational Complexity of VS

• The S set for conjunctive feature vectors and tree-

structured attributes is linear in the number of features
and the number of training examples.
• The G set for conjunctive feature vectors and tree-
structured attributes can be exponential in the number
of training examples.
• In more expressive languages, both S and G can grow
exponentially.
• The order in which examples are processed can
significantly affect computational complexity.

40
Exponential size of G
• n Boolean attributes
• 1 positive example: (T, T, .., T)
• n/2 negative examples:
– (F,F,T,..T)
– (T,T,F,F,T..T)
– (T,T,T,T,F,F,T..T)
– ..
– (T,..T,F,F)
• Every hypothesis in G needs to choose from n/2 2-element
sets.
– Number of hypotheses = 2n/2

41
Summary
• Concept learning as search through H
• General-to-specific ordering over H
• Version space candidate elimination algorithm
• S and G boundaries characterize learner’s uncertainty
• Learner can generate useful queries
• Inductive leaps possible only if learner is biased!
• Inductive learners can be modeled as equiv deductive
systems
• Biggest problem is inability to handle data with errors
– Overcome with procedures for learning decision trees

42