Introduction to Artificial Intelligence

(Comp551)
1

JIMMA UNIVERSITY
JIMMA INSTITUTE OF TECHNOLOGY
DEPARTMENT OF COMPUTING

CHAPTER NIGN
LEARNING FROM OBSERVATIONS
 Topics we will cover
2

Learning agents
Inductive learning
Decision tree learning
 Learning
3

Learning is essential for unknown environments,

 i.e., when designer lacks omniscience.
 Omniscience (all-knowing with infinite knowledge).
 An agent is autonomous if its behavior is determined by its own
experience.
Learning is useful as a system construction
method,
 i.e., expose the agent to reality rather than trying to
write it down.
Learning modifies the agent's decision
mechanisms to improve performance.
Learning agents
4

Learning agents
A learning agent can be divided into
four conceptual components:
 Performance Element can be
replaced with any of the 4
agent types .
 The Learning Element is
responsible for suggesting
improvements to any part of the
performance element.
 The input to the learning element
comes from the Critic.
 The Problem Generator is
responsible for suggesting
Figure - The basic structure of a learning
actions that will result in new agent
knowledge about the world being
acquired.
 Learning element
5

 Design of a learning element is affected by:

 Which components of the performance element should be modified?
 What feedback is available to guide the learning process?
 How is the performance element represented?
 Type of feedback:
 Depending on the answer to this question, we can define three
types of learning process:
 Supervised learning: correct answers for each example.
• The agent learns a function from examples of its inputs and
outputs.
 Unsupervised learning: correct answers not given.
• The agent learns patterns in the inputs, without actually
knowing what the correct output should be.
 Reinforcement learning: occasional rewards.
• The outputs of the function are not explicitly specified, but the
agent may get an occasional reward for doing good things.
 Inductive learning
6

 The simplest type of learning: inductive learning.

 Simplest form: learn a function from examples.

 f is the target function.

 An example is a pair (x, f(x))

Problem: find a hypothesis h

such that h ≈ f (h is an approximation of f)
given a training set of examples

 This is a highly simplified model of real learning:

 Ignores prior knowledge.
 Assumes examples are given.
 Inductive learning method (1)
7

 Construct/adjust h to agree with f on training set.

 (h is consistent if it agrees with f on all examples)
 E.g., curve fitting:
 Inductive learning method (2)
8

 Construct/adjust h to agree with f on training set.

 (h is consistent if it agrees with f on all examples)
 E.g., curve fitting:
 Inductive learning method (3)
9

 Construct/adjust h to agree with f on training set.

 (h is consistent if it agrees with f on all examples)
 E.g., curve fitting:
 Inductive learning method (4)
10

 Construct/adjust h to agree with f on training set.

 (h is consistent if it agrees with f on all examples)
 E.g., curve fitting:
 Inductive learning method (5)
11

 Construct/adjust h to agree with f on training set.

 (h is consistent if it agrees with f on all examples)
 E.g., curve fitting:
 Inductive learning method (6)
12
 Construct/adjust h to agree with f on training set.
 (h is consistent if it agrees with f on all examples)
 E.g., curve fitting:

 Ockham’s razor: prefer the simplest hypothesis

consistent with data.
 Learning decision trees
13

 This is one of the simplest and most successful of learning algorithms:

decision trees.
 Problem: Suppose we arrive at a restaurant intending to have dinner. If
there is no space, depending on certain circumstances we may decide
whether to wait for a table at a restaurant, based on the following
attributes:
1. Alternate: is there an alternative restaurant nearby?
2. Bar: is there a comfortable bar area to wait in?
3. Fri/Sat: is today Friday or Saturday?
4. Hungry: are we hungry?
5. Patrons: number of people in the restaurant (None, Some, Full)
6. Price: price range ($, $$, $$$)
7. Raining: is it raining outside?
8. Reservation: have we made a reservation?
9. Type: kind of restaurant (French, Italian, Thai, Burger)
10. WaitEstimate: estimated waiting time (0-10, 10-30, 30-60, >60)
 Attribute-based representations
14

 Examples described by attribute values.

 E.g., situations where I will/won't wait for a table:
o (Decision to wait or not to wait a table at a restaurant)

Table– Training set for the restaurant learning problem

 Classification of example is either positive (T) or negative (F)
 Decision trees
15
 One possible representation for hypotheses.
 E.g., here is the “true” tree for deciding whether to wait:
 Decision to wait or not to wait a table at a restaurant.

Figure – A decision tree for the restaurant problem

 Decision Trees: Expressiveness
16

 Decision trees can express any function of the input attributes.

 E.g., for Boolean functions, truth table row → path to leaf:

 Trivially, there is a consistent decision tree for any training set with one path to
leaf for each example but it probably won't generalise to new examples.

 Prefer to find more compact decision trees.

 Hypothesis Space of Decision Trees
17

How many distinct decision trees with n Boolean attributes?

 With decision trees, if we have n attributes, there are 2 to the power 2n
different possible decision trees.
= number of Boolean functions
n
= number of distinct truth tables with 2n rows = 22

 E.g., with 6 Boolean attributes, there are 18,446,744,073,709,551,616

trees
 Performance measurement
18
 How do we know that h ≈ f (h is an approximation of f) ?
1. Use theorems of computational/statistical learning theory.
2. Try h on a new test set of examples.
 Learning curve = % correct on test set as a function of training set size

Figure – A sample learning curve

 Summary
19

Learning needed for unknown environments, lazy

designers.
Learning agent = performance element + learning
element.
For supervised learning, the aim is to find a
simple hypothesis approximately consistent with
training examples.
Decision tree learning using information gain .
Learning performance = prediction accuracy
measured on test set.
 Exercise:
20

 Below is a training set for the problem of determining whether England

will win the football world cup. There are three input attributes (one
Boolean and one with three discrete values, and a Boolean outcome)
Using the information theory approach to selecting attributes,
construct a decision tree to represent this function.

INPUT ATTRIBUTES OUTPUT

RooneyPlays? Temperature? EnglandWin?
Yes 20-30oC Yes
Yes <20oC Yes
No <20oC No
No 20-30oC No
Yes >30oC No
No >30oC No
 Exercise Solution
21
 Therefore the maximum information gain comes from choosing the RooneyPlays
attribute. We make this the root of the tree. Now we have the following tree:

 We can see that the negative branch requires no further classification, as all 3 outcomes
are Lose. The positive branch requires further classification, so we use the Temperature
attribute, resulting in the following final decision tree: