8/12/2019 AI - Agents

1/31

LECTURE 2

INTELLIGENT AGENTS

8/12/2019 AI - Agents

2/31

Contents

What is an Agent?Agents and ObjectsEnvironments

An Abstract Architecture for AgentsReactive AgentsAgents with Internal States

Goal-based AgentsUtility-based Agents

8/12/2019 AI - Agents

3/31

What is an Agent?The main point about agents is they areautonomous : capable of acting independently,

exhibiting control over their internal state.Thus: an agent is a computer system capableof autonomous action in some environment in

order to meet its design objectives.

AGENT

ENVIRONMENT

input output

8/12/2019 AI - Agents

4/31

What is an Agent?Trivial (non-interesting) agents:

thermostat

UNIX daemon (e.g., biff)An intelligent agent is a computer systemcapable of flexible autonomous action in

some environment.By flexible , we mean:

reactive

pro-activesocial

8/12/2019 AI - Agents

5/31

ReactivityIf a programs environment is guaranteed to be fixed, theprogram need never worry about its own success orfailure program just executes blindly.

Example of fixed environment: compilerThe real world is not like that: things change, informationis incomplete. Many (most?) interesting environments are

dynamic.Software is hard to build for dynamic domains: programmust take into account possibility of failure ask itselfwhether it is worth executing!A reactive system is one that maintains an ongoinginteraction with its environment, and responds to changesthat occur in it (in time for the response to be useful).

8/12/2019 AI - Agents

6/31

ProactivenessReacting to an environment is easy (e.g.,stimulus response rules).But we generally want agents to do thingsfor us.

Hence goal directed behavior.Pro-activeness = generating andattempting to achieve goals; not driven

solely by events; taking the initiative.Recognizing opportunities.

8/12/2019 AI - Agents

7/31

Reactive and Proactive Behavior

We want our agents to be reactive,responding to changing conditions in anappropriate (timely) fashion.We want our agents to systematically work

towards long-term goals.These two considerations can be at odds withone another.

Designing an agent that can balance the tworemains an open research problem.

8/12/2019 AI - Agents

8/31

8/12/2019 AI - Agents

9/31

Other PropertiesOther properties, sometimes discussed in thecontext of agency:

mobility : the ability of an agent to move around anelectronic networkveracity : an agent will not knowingly communicate falseinformation

benevolence : agents do not have conflicting goals, andthat every agent will therefore always try to do what isasked of itrationality : agent will act in order to achieve its goals, andwill not act in such a way as to prevent its goals beingachieved at least insofar as its beliefs permitlearning/adaption : agents improve performance over time

8/12/2019 AI - Agents

10/31

Contents

What is an Agent?

Agents and ObjectsEnvironments

An Abstract Architecture for AgentsReactive AgentsAgents with Internal States

Goal-based AgentsUtility-based Agents

8/12/2019 AI - Agents

11/31

Agents and Objects

Are agents just objects by another name?

Object:encapsulates some statecommunicates via message passing

has methods, corresponding to operationsthat may be performed on this state

8/12/2019 AI - Agents

12/31

Agents and Objectsagents are autonomous

agents embody stronger notion of autonomy thanobjects, and in particular, they decide for themselveswhether or not to perform an action on request fromanother agent

agents are smart

capable of flexible (reactive, pro-active, social) behavior,and the standard object model has nothing to say aboutsuch types of behavior

agents are active a multi-agent system is inherently multi-threaded, in thateach agent is assumed to have at least one thread ofactive control

8/12/2019 AI - Agents

13/31

8/12/2019 AI - Agents

14/31

Contents

What is an Agent?

Agents and ObjectsEnvironments

An Abstract Architecture for AgentsReactive AgentsAgents with Internal States

Goal-based AgentsUtility-based Agents

8/12/2019 AI - Agents

15/31

Accessible / Inaccessible

An accessible environment is one in which

the agent can obtain complete, accurate,up-to-date information about theenvironments state.

Most moderately complex environments(including, for example, the everydayphysical world and the Internet) areinaccessible.The more accessible an environment is, thesimpler it is to build agents to operate in it.

8/12/2019 AI - Agents

16/31

Deterministic / Non-deterministic

A deterministic environment is one in which

any action has a single guaranteed effect there is no uncertainty about the state thatwill result from performing an action.

The physical world can to all intents andpurposes be regarded as non-deterministicNon-deterministic environments presentgreater problems for the agent designer.

8/12/2019 AI - Agents

17/31

Episodic / Non-episodic

In an episodic environment, the performance of an

agent is dependent on a number of discreteepisodes, with no link between the performance ofan agent in different scenarios.

Episodic environments are simpler from the agentdevelopers perspective because the agent candecide what action to perform based only on thecurrent episode it need not reason about theinteractions between this and future episodes.

8/12/2019 AI - Agents

18/31

Static / Dynamic

A static environment is one that can be assumed to

remain unchanged except by the performance ofactions by the agent.A dynamic environment is one that has other

processes operating on it, and which hence changesin ways beyond the agents control.Other processes can interfere with the agentsactions (as in concurrent systems theory).The physical world is a highly dynamic environment.

8/12/2019 AI - Agents

19/31

Discrete / Continuous

An environment is discrete if there are a fixed, finite

number of actions and percepts in it.Russell and Norvig give a chess game as anexample of a discrete environment, and taxi driving

as an example of a continuous one.Continuous environments have a certain level ofmismatch with computer systems.

Discrete environments could in principle be handledby a kind of lookup table.

8/12/2019 AI - Agents

20/31

8/12/2019 AI - Agents

21/31

Contents

What is an Agent?

Agents and ObjectsEnvironments

An Abstract Architecture for AgentsReactive AgentsAgents with Internal States

Goal-based AgentsUtility-based Agents

8/12/2019 AI - Agents

22/31

Reactive Agents

8/12/2019 AI - Agents

23/31

Reactive Agents

Percepts from the sensors map directly to

actions.Condition-action rules defines the mapping.Can be very efficient.Complex decisions are possible, but:

Constructing the rules is a difficult task.

Any change in the environment may make someor all the rules invalid.

8/12/2019 AI - Agents

24/31

Agents with Internal States

8/12/2019 AI - Agents

25/31

Agents with Internal States

If full percept sequence is available, reactive

agents can perform complex tasks.However, the condition-action rules are

numerous and complex.More efficient to use internal states:

Need to understand how the world evolves.

Need to understand the effects of actions.This is a big increase in complexity.

8/12/2019 AI - Agents

26/31

Goal-based Agents

8/12/2019 AI - Agents

27/31

Goal-based Agents

Understanding the environment is often not

enough.What is the agent trying to achieve?

Proactive behaviour

Goal-based agents separate mechanicalactions from the overall intention of the agent.

8/12/2019 AI - Agents

28/31

Utility-based Agents

8/12/2019 AI - Agents

29/31

Utility-based Agents

Goals are crude distinctions but the agent

may have to choose many paths all of whichsatisfy the goal.The utility of each path should determine theaction of the agent.For example, when playing chess, our goal isto win the game but we must decide onindividual moves on the basis of the utility ofthe board position.

8/12/2019 AI - Agents

30/31

Reading

Agent or Program , Franklin and Graesser

8/12/2019 AI - Agents

31/31