Knowledge is the information about a domain that can be used to solve problems in that domain. To solve many problems requires much knowledge, and this knowledge must be represented in the computer. As part of designing a program to solve problems, we must define how the knowledge will be represented.

1. DEPARTMENT OF CS &IT
KNOWLEDGE IN LEARNING
PRESENTED BY:
S.SABTHAMI
I.MSC(IT)
Nadar saraswathi college of arts and science

2. A logical formulation of learning
 What’re Goal and Hypotheses
 Goal predicate Q - WillWait
 Learning is to find an equivalent logical
expression we can classify examples
 Each hypothesis proposes such an
expression - a candidate definition of Q
 r WillWait(r) Pat(r,Some) 
Pat(r,Full) Hungry(r)Type(r,French) 

3. A logical formulation of learning
 Hypothesis space is the set of all hypotheses
the learning algorithm is designed to
entertain.
 One of the hypotheses is correct:
H1 V H2 V…V Hn
 Each Hi predicts a certain set of examples -
the extension of the goal predicate.
 Two hypotheses with different extensions
are logically inconsistent with each other,
otherwise, they are logically equivalent.

4. Examples
 An example is an object of some logical
description to which the goal concept may
or may not apply.
 Alt(X1)^!Bar(X1)^!Fri/Sat(X1)^…
 Ideally, we want to find a hypothesis that
agrees with all the examples.
 The relation between f and h are: ++, --, +-
(false negative), -+ (false positive). If the
last two occur, example I and h are logically
inconsistent.

5. Current-best hypothesis search
 Maintain a single hypothesis
 Adjust it as new examples arrive to maintain
consistency
Generalization
Specialization

6. Example of WillWait
 Problems: nondeterministic, no
guarantee for simplest and
correct h, need backtrack

7. Least-commitment search
 Keeping only one h as its best guess is the
problem -> Can we keep as many as
possible?
 Version space (candidate elimination)
Algorithm
 incremental
 least-commitment

8. Least-commitment search
 From intervals to boundary sets
 G-set and S-set
 S0 – the most specific set contains nothing <0,0,…,0>
 G0 – the most general set covers everything <?,?,…,?>
 Everything between is guaranteed to be
consistent with examples.
 VS tries to generalize S0 and specialize G0
incrementally

9. Version space
 Generalization and specialization
 find d-sets that contain only true/+, and true/-;
 Sj can only be generalized and Gj can only be specialized
 False positive for Si, too general, discard it
 False negative for Si, too specific, generalize it minimally
 False positive for Gi, too general, specialize it minimally
 False negative for Gi, too specific, discard it

10. Version space
 When to stop
 One concept left (Si = Gi)
 The version space collapses (G is more special than S, or..)
 Run out of examples
 An example with 4 instances from Tom Mitchell’s
book
 One major problem: can’t handle noise

11. Using prior knowledge
 For DT and logical description learning, we
assume no prior knowledge
 We do have some prior knowledge, so how
can we use it?
 We need a logical formulation as opposed to
the function learning.

12. Inductive learning in the logical setting
 The objective is to find a hypothesis that
explains the classifications of the examples,
given their descriptions.
Hypothesis ^ Description |= Classifications
 Hypothesis is unknown, explains the
observations
 Descriptions - the conjunction of all the example
descriptions
 Classifications - the conjunction of all the
example classifications
 Knowledge free learning
 Decision trees
 Description = Classifications

13. A procecumulative learning ss
 Observations, K-based learning,
Hypotheses, and prior knowledge
 The new approach is to design agents that
already know something and are trying to
learn some more.
 Intuitively, this should be faster and better
than without using knowledge, assuming
what’s known is always correct.

14. Some examples of using knowledge
 One can leap to general conclusions after
only one observation.
 Your such experience?
 Traveling to Brazil: Language and name
 A pharmacologically ignorant but
diagnostically sophisticated medical
student …

15. Some general schemes
 Explanation-based learning (EBL)
 Hypothesis^Description |= Classifications
 Background |= Hypothesis
 doesn’t learn anything factually new from instance
 Relevance-based learning (RBL)
 Hypothesis^Descriptions |= Classifications
 Background^Descrip’s^Class |= Hypothesis
 deductive in nature
 Knowledge-based inductive learning (KBIL)
 Background^Hypothesis^Descrip’s |=
Classifications

16. Inductive logical programming (ILP)
 ILP can formulate hypotheses in general
first-order logic
 Others like DT are more restricted languages
 Prior knowledge is used to reduce the
complexity of learning:
 prior knowledge further reduces the H space
 prior knowledge helps find the shorter H
 Again, assuming prior knowledge is correct

17. Explanation-based learning
 A method to extract general rules from individual
observations
 The goal is to solve a similar problem faster next
time.
 Memoization - speed up by saving results and
avoiding solving a problem from scratch
 EBL does it one step further - from observations to
rules

18. Basic EBL
 Given an example, construct a proof tree using the
background knowledge
 In parallel, construct a generalized proof tree for
the variabilized goal
 Construct a new rule (leaves => the root)
 Drop any conditions that are true regardless of the
variables in the goal

19. Efficiency of EBL
 Choosing a general rule
 too many rules -> slow inference
 aim for gain - significant increase in speed
 as general as possible
 Operationality - A subgoal is operational means it is
easy to solve
 Trade-off between Operationality and Generality
 Empirical analysis of efficiency in EBL

20. Learning using relevant information
 Prior knowledge: People in a country
usually speak the same language
Nat(x,n) ^Nat(y,n)^Lang(x,l)=>Lang(y,l)
 Observation: Given nationality, language is
fully determined
 Given Fernando is Brazilian & speaks Portuguese
Nat(Fernando,B) ^ Lang(Fernando,P)
 We can logically conclude
Nat(y,B) => Lang(y,P)

21. Functional dependencies
 We have seen a form of relevance:
determination - language (Portuguese) is a
function of nationality (Brazil)
 Determination is really a relationship
between the predicates
 The corresponding generalization follows
logically from the determinations and
descriptions.

22. Functional dependencies
 Determinations specify a sufficient basis
vocabulary from which to construct hypotheses
concerning the target predicate.
 A reduction in the H space size should make it
easier to learn the target predicate