2

BLG521E - Lecture 2

Outline • • • • •

BLG521E  Advanced Artificial Intelligence Lecture 2: Intelligent Agents

Agents and Environments Rationality PEAS Environment Types Agent Types

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Agents

2

Agents • Agents perceive their own actions  – Effects?

Percept: the agen’t perceptual input Percept sequence: the complete history A i Action choices depend on the percept sequence h i d d h Agent function, abstract mathematical description  (agent’s behavior) • Agent program implements the function

• • • •

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

3

Vacuum‐Cleaner World 

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

4

A Vacuum‐Cleaner Agent Function

5

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

6

1

BLG521E - Lecture 2

A Vacuum‐Cleaner Agent Function

Vacuum Cleaner Agent ‐ PM  • The amount of dirt cleaned up in a single eight‐hour  shift. • Rewarding agent for having a clean floor. • Factoring amount of electricity consumed and the  amount of noise generated amount of noise generated    • Design PM according to what you want

What is the right way to fill out the table?

What makes an agent good, bad, intelligent or stupid? Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

7

Rationality

8

Rational Agent

• A rational agent does the right thing • What is rational at any given time depends on:

• For each possible percept sequence, a rational agent should select an action that is expected to maximize its performance efficiency, given the evidence provided by the percept sequence and whatever built‐in knowledge the agent has. g g

– The performance measure that defines the criterion of  success – The agent The agent’ss prior knowledge of the environment prior knowledge of the environment – The actions that the agent can perform – The agent’s percept sequence to date

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

• With a rational agent, what you ask is what you get

9

Rationality vs. Perfection

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

10

Further Dimensions in Rationality

• Omniscience is impossible in reality • Agents don’t estimate the actual outcome of actions

• Information gathering

• Rationality maximizes expected outcome, while  perfection maximizes actual performance f ti i i t l f

• Learning

– Exploration – Helps maximize the expected outcome

• Autonomy • With or without initial knowledge

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

11

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

12

2

BLG521E - Lecture 2

Agents  Environments

The Nature of Environments

• Task environment forms the problem

• PEAS for task environments:

– Rational agents are the solutions

– – – –

• The task environment affects the appropriate design  of the agent of the agent

Performance measure  Environment Actuators Sensors

• PEAS for automated taxi driver

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

13

Properties of Task Environments

14

Examples of Task Environments

• Fully observable vs. Partially observable • Deterministic vs. Stochastic

Task Environment

– strategic

Observable

Deterministic

Episodic

Static

Discrete

Agents

Crossword  puzzle

• Episodic vs. Sequential • Static vs. Dynamic S i D i

Chess with a  clock Poker

– semidynamic

Backgammon

• Discrete vs. Continous • Single agent vs. Multiagent

Taxi driving

– competitive – cooperative Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Examples of Task Environments Task Environment

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

15

Examples of Task Environments

Observable

Deterministic

Episodic

Static

Discrete

Agents

Task Environment

Observable

Deterministic

Episodic

Static

Discrete

Agents

Fully

Deterministic

Sequential

Static

Discrete

Single

Crossword  puzzle

Fully

Deterministic

Sequential

Static

Discrete

Single

Chess with a  clock

Chess with a  clock

Fully

Strategic

Sequential

Semi

Discrete

Multi

Poker

Poker

Backgammon

Backgammon

Taxi driving

Taxi driving

Crossword  puzzle

16

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

17

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

18

3

BLG521E - Lecture 2

Examples of Task Environments Task Environment

Examples of Task Environments

Observable

Deterministic

Episodic

Static

Discrete

Agents

Task Environment

Crossword  puzzle

Fully

Deterministic

Sequential

Static

Discrete

Single

Chess with a  clock

Fully

Strategic

Sequential

Semi

Discrete

Multi

Partially

Stochastic

Sequential

Static

Discrete

Multi

Poker

Poker

Observable

Deterministic

Episodic

Static

Discrete

Agents

Crossword  puzzle

Fully

Deterministic

Sequential

Static

Discrete

Single

Chess with a  clock

Fully

Strategic

Sequential

Semi

Discrete

Multi

Backgammon

Backgammon

Taxi driving

Taxi driving

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

19

Examples of Task Environments

Partially

Stochastic

Sequential

Static

Discrete

Multi

Fully

Stochastic

Sequential

Static

Discrete

Multi

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

20

The Structure of Agents • The job of AI is to design the agent program

Task Environment

Observable

Deterministic

Episodic

Static

Discrete

Agents

Crossword  puzzle

Fully

Deterministic

Sequential

Static

Discrete

Single

Chess with a  clock

Fully

Strategic

Sequential

Semi

Discrete

Multi

Partially

Stochastic

Sequential

Static

Discrete

Multi

Fully

Stochastic

Sequential

Static

Discrete

Multi

Partially

Stochastic

Sequential

Dynamic

Continuous

Multi

Poker Backgammon Taxi driving

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

• Agent architecture • Agent = Architecture + Program

21

Agent Types • • • •

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

22

Simple Reflex Agents

Simple reflex agents Model‐based reflex agents Goal‐based agents Utility‐based agents

• All these agents can be converted into learning  agents

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

23

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

24

4

BLG521E - Lecture 2

Reflex Vacuum Agent Program

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Simple‐Reflex Agent Program

25

Model‐based Reflex Agents

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

26

Model‐based Agent Program

27

Goal‐Based Agents

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

28

Utility‐based Agents

29

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

30

5

BLG521E - Lecture 2

Learning Agents

Advanced Artificial Intelligence (BLG521E) @ ITU:: Computer Engineering Department, Dr. Sanem Sarıel‐Talay

31

6