Chapter 8

1. King Saud University College of Computer and Information Sciences Information Technology Department IT422 - Intelligent systems Chapter 8 Machine Learning

2. Introduction • What is learning? • Learning in humans consists of (at least): • memorization, comprehension, learning from examples. • Learning from examples • Square numbers: 1, 4, 9 ,16 • 1 = 1 * 1; 4 = 2 * 2; 9 = 3 * 3; 16 = 4 * 4; • What is next in the series? • We can learn this by example quite easily

3. Introduction • What is learning? “Learning denotes changes in a system that enable the system to do the same task more efficiently next time”. (Hubert Simon, 1983) • An agent is learning if it improves its performance on future tasks after making observations about the world.

4. Introduction • What is learning? • "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 E". (Mitchell, 1997) • Given: a task T, a performance measure P, some experience E with the task. • Goal: generalize the experience in a way that allows to improve the performance on the task.

5. Why would we want an agent to learn? • The designer can not anticipate all situations in which the agent may be. • For example, a robot navigating a maze, robot in space. • The designer can not anticipate all changes over time. • For example, stock market prediction. • Sometimes the designers have no idea how to program the solutions themselves (unknown function). • For example: face recognition.

6. Components to be learned • Design of a learning element is affected by • Which component is to be improved • What prior knowledge the agent already has. • What feedback is available to learn from. • What representation is used for the data and the component

7. Components to be learned Consider an agent training to become a taxi driver • When the instructor shouts “Brake!” the agent learn a condition–action rule for when to brake; also when the instructor does not shout . • By seeing many camera images that it is told contain buses, it can learn to recognize them. • By trying actions and observing the results. Ex. braking hard on a wet road. • When it receives no tip from passengers, it can learn a useful component who have been shaken up during the trip of its overall utility function.

8. Types of Learning • In order to learn, the agent needs to observe the world  feedback. • The different types of feedback determine the different types of learning: • Supervised learning • Unsupervised learning • Semi-supervised learning • Reinforcement learning

9. Types of Learning • Supervised learning: The agent observes a set of input-output examples (labeled examples) and learns a map from inputs to outputs. • Classification (Categorization): output is discrete . Learn why certain objects are categorized a certain way. E.g.: spam email, why are dogs, cats and humans mammals, but trout, mackerel and tuna are fish? • Binary classification (Boolean): there are only two values. • Regression(Prediction): output is real-valued . Learn how to predict how to categorize unseen objects E.g., Given examples of financial stocks and a categorization of them into safe and unsafe stocks Learn how to predict whether a new stock will be safe. • Unsupervised learning: No explicit feedback is given, only the inputs (unlabeled examples). The agent learns patterns in the input. • Ex. “good traffic days” • Semi-supervised learning: The agent is given some labeled examples (generally a few) and some unlabeled examples and tries to learn a mapping. • Reinforcement learning: The agent learns from a series of rewards and punishments, and based on these adapts its behavior (e.g. playing chess) .

10. Supervised Learning • Given a training set of N example input-output pairs: (x1, y1), (x2, y2), … (xN, yN), where, yj = f(xj), where f is unknown function, the goal is to find a function h that approximates f. • The function h is called a hypothesis. • How to measure the accuracy of h? • We give a test set of examples, which is different from the training set. • The hypothesis generalizes well if it correctly predicts the output for the test set.

11. How to select a hypothesis (a) (b) First, select the hypothesis space: in this case, the set of polynomials. (a): The line is consistent with the data. (b): The high-degree polynomial is also consistent with the data. Ockham’s razor: Choose the simplest hypothesis which is consistent with the data.

12. Decision Trees • A decision tree represents a function that has multiple inputs but a single output a “decision”. • We focus on discrete input and Boolean output (Boolean classification) • A decision tree reaches the decision by a set of tests on the attributes (the inputs). Thus, the internal nodes are the tests and the leaf nodes are the decisions. • Example: Decision nodes Test nodes

13. Decision Trees • A more complex example: deciding to wait at a restaurant: • The attributes : • Alternate: whether there is a suitable alternative restaurant nearby. • Bar: whether the restaurant has a comfortable bar area to wait in. • Fri I Sat: true on Fridays and Saturdays. • Hungry: whether we are hungry. • Patrons: how many people are in the restaurant (values are None, Some, and Full). • Price: the restaurant's price range ($, $$, $$$). • Raining: whether it is raining outside. • Reservation: whether we made a reservation. • Type: the kind of restaurant (French, Italian, Thai, or burger). • WaitEstimate: the wait estimated by the host (0-10 minutes, 10-30, 30-60, or >60).

14. Decision Trees • Classification of examples is positive (T) or negative (F)

15. Decision Trees • This is the real function. • Our goal is to learn this function from examples.

16. Decision Trees • A decision tree can be expressed as propositional logic sentence (Boolean function) in DNF (disjunctive normal form): • Goal  (Path1 V Path2 V … Pathn), where Pathi= (Attribute1 = Valuek1  Attribute2 = Valuek2 …) • The same Boolean function can have many representations as a decision tree (just change the order of the attributes)  We want the smallest possible tree: • Example: The decision tree of P  (Q  R)

17. Decision Trees The order of the attributes: Q, R,P The order of the attributes: P, Q, R Smaller number of nodes  The order is important A decision tree for the function: P  (Q  R).

18. Decision Trees • For n (Boolean) attributes there are 2^(2^n) different Boolean functions, and the number of decision trees is much larger (more than n! 2^(2^n) ) • Example: n = 6, there are approximately 18.4 x 10^18 possible Boolean functions • Exhaustive search is impossible in practice  Learning the decision tree  greedy heuristic search • How to choose the most important attribute and build the decision tree? • Several algorithms exist. Ex. ID3 (Iterative Dichotomiser 3)

19. Summary • Learning takes many forms, depending on the nature of the agent, the component to be improved, and the available feedback. • Learning can be supervised, unsupervised, semi-supervised learning, and reinforcement learning, depending on the given feedback. • Decision trees are powerful tools for classification, they can represent rules in tree structure where each node is either test or decision node.