Artificial Intelligence Programming: Decision Trees and Learning Agents - Prof. Christophe, Study notes of Computer Science

Download Artificial Intelligence Programming: Decision Trees and Learning Agents - Prof. Christophe and more Study notes Computer Science in PDF only on Docsity!

Artificial Intelligence Programming

Decision Trees^ Chris BrooksDepartment of Computer ScienceUniversity of San Francisco

Department of Computer Science — University of San Francisco – p. 1/

Rule Learning • Previously, we’ve assumed that background knowledge wasgiven to us by experts. ◦^ Focused on how to use that knowledge. • Today, we’ll talk about how to acquire that knowledge fromobservation. • Focus on learning propositional rules ◦^ sunny

∧^ warm

→^ P layT ennis

◦^ cool

∧^ (rain

∨^ strongW ind

)^ → ¬

P layT ennis

Department of Computer Science — University of San Francisco – p. 2/

Learning • What does it mean for an agent to learn? • Agent acquires new knowledge • Agent changes its behavior • Agent improves its performance measure on a given task

Department of Computer Science — University of San Francisco – p. 4/

Learning Agents • Recall that at the beginning of the semester we talked about learning agents^ Performance standard^ Agent

Environment

Sensors Performanceelement changesknowledge Critic feedback Learningelementlearninggoals Problemgenerator

Actuators

Department of Computer Science — University of San Francisco – p. 5/

Deduction vs. Induction • Up to now, we’ve looked at cases where our agent is givengeneral knowledge and uses this to solve a particular problem. ◦^ It always takes Edge 2 minutes to cross the river, Suckalways cleans a room, etc. ◦^ This general-to-specific reasoning is known as

deduction

◦^ Advantage: deduction is sound, assuming your knowledgeis correct. • Sometimes, you may not have general information about aproblem. • Instead, you might have

data

about particular instances of a

problem. • the problem then is to figure out a general rule from specificdata. • This is called

induction

• most learning is an inductive process.

◦^ Problem: induction is not sound.

Department of Computer Science — University of San Francisco – p. 7/

Example • We’ll begin with the example of an agent deciding whether weshould play tennis on a given day. • There are four observable percepts: ◦^ Outlook (sunny, rainy, overcast) ◦^ Temperature (hot, mild, cool) ◦^ Humidity (high, low) ◦^ Wind (strong, weak) • We don’t have a model, but we do have some data about pastdecsions. • Can we induce a general rule for when to play tennis?

Department of Computer Science — University of San Francisco – p. 8/

Types of Learning Tasks • Unsupervised Learning ◦^ In this case, there is no teacher to provide examples. ◦^ The agent typically tries to find a “concept” or pattern indata. ◦^ Statistical methods such as clustering fall into this category ◦^ Our agent might be told that day1, day 4 and day 7 aresimilar and need to determine what characteristics makethese days alike.

Department of Computer Science — University of San Francisco – p. 10/

Types of Learning Tasks • Reinforcement Learning ◦^ This is a particular version of learning in which the agentonly receives a

reward

for taking an action.

◦^ May not know how optimal a reward is. ◦^ Will not know the “best” action to take ◦^ Our agent might be presented with a Sunny, Hot, Lowhumidity, Strong wind day and asked to choose whether toplay tennis. ◦^ It chooses ’yes’ and gets a reward of 0.

Department of Computer Science — University of San Francisco – p. 11/

Classification • the particular learning problem we are focusing on issometimes known as

classification

◦^ For a given input, determine which class it belongs to. • Programs that can perform this task are referred to as classifiers

Department of Computer Science — University of San Francisco – p. 13/

Defining the Learning Problem • We can phrase the learning problem as that of estimating afunction

f^ that tells us how to classify a set of inputs.

-^ An example is a set of inputs

x^ and the corresponding

f^ (x)^

the class that

x^ belongs to. ◦^ << Overcast, Cool, Low, W eak >, playT ennis > • We can define the learning task as follows: ◦^ Given a collection of examples of

f^ , find a function

H^ that

approximates

f^ for our examples.

◦^ H^ is called a

hypothesis

Department of Computer Science — University of San Francisco – p. 14/

Inductive Bias • Notice that induction is not sound. • In picking a hypothesis, we make an educated guess. • the way in which we make this guess is called a

bias.

-^ All learning lagorithms have a bias; identifying it can help youunderstand the sorts of errors it will make. •^ Examples:^ ◦^

Occam’s razor ◦ Most specific hypothesis. ◦ Most general hypothesis. ◦ Linear function

Department of Computer Science — University of San Francisco – p. 16/

Observing Data • Agents may have different means of observing examples of ahypothesis. • A batch

learning algorithm is presented with a large set of data all at once and selects a single hypothesis. • An^ incremental

learning algorithm receives examples one at a time and continually modifies its hypothesis.^ ◦^ Batch is typically more accurate, but incremental may fitbetter with the agent’s environment. • An^ active

learning agent is able to choose examples.

-^ A^ passive

learning agent has examples presented to it by an outside source.^ ◦^ Active learning is more powerful, but may not fit with theconstraints of the domain.

Department of Computer Science — University of San Francisco – p. 17/

Learning Decision Trees • Decision trees are data structures that provide an agent with ameans of classifying examples. • At each node in the tree, an attribute is tested.

Feathers? Yes^

No

Does it Fly?^

Does it Eat Meat?

No Yes Yes Does it havelong legs?^

Albatross Yes Ostrich No No Penguin

Does it haveBlack Stripes?Yes

No Tiger

Does it havea long neck? Cheetah

Yes^

No Giraffe^

Zebra Department of Computer Science — University of San Francisco – p. 19/

Another Example • R & N show a decision tree fordetermining whether to wait at a busyrestaurant. • The problem has the followinginputs/attributes: ◦^ Alternative nearby ◦^ Has a bar ◦^ Day of week ◦^ Hungriness ◦^ Crowd ◦^ Price ◦^ Raining? ◦^ Reservation ◦^ Type of restuarant ◦^ Wait estimate

No^ Yes No Yes

No^ Yes^ No^

Yes

None^ Some

Full >60 30-^

10-^

0- No^ Yes

Hungry?^ Alternate? Reservation? Bar?^

Raining?

Patrons?^ Alternate?^ Fri/Sat? No^ Yes^ No^

Yes Yes^ Yes^ No

Yes No Yes Yes No Yes WaitEstimate? No Yes Yes No • Note that

not^

all^ atributes

are used.

Department of Computer Science — University of San Francisco – p. 20/