Intelligent Agents Devika Subramanian Comp440 Lecture 1.

Intelligent Agents

Devika SubramanianComp440Lecture 1

Intelligent Agents

An agent is anything that can be viewed as perceiving its environment through its sensors and acting uponthat environment through its effectors.

environment

agent

sensors

effectors

Objective: design agents that perform well in their environments

Informal agent descriptions Thermostat

Percepts: temperature sensor Actions: open/close valve Performance measure: maintaining user-set

temperature Environment: room/house

Internet Newsweeder Percepts: words, bitmaps Actions: word vector counts, cosine transforms, etc Performance measure: retrieving relevant news posts Environment: Internet newsgroups

Specifying performance measures

How do we measure how well an agent is doing? External performance measure or

self-evaluation? When do we measure performance

of agent? Continuous, periodic or one-shot

evaluation?

Specifying performance measures formally Performance measures are external. The environment provides feedback to

the agent in the form of a function mapping the environment’s state history to a real number.

Performance feedback can be provided after each move, periodically, or at the very end.

*: SV

Example of performance measure for thermostat

s0s1

sn-1 sn

Goal: maximizewith discount factor

The ambient temperature is being sampledat periodic intervals.

otherwise 100-

DESIRED )re( temperatuif 100)(

tt ssV

...)(...)()( 10 nn sVsVsV

10

Statesof theenvironment

Ideal rational agent

An ideal rational agent performs actions that are expected to maximize its performance measure on the basis of the evidence provided by its percept sequence and whatever built-in knowledge the agent has. An ideal rational agent is not omniscient. Doing actions to gather information is part of

rational behavior. Rationality of agents judged using performance

measure, percept sequence, agent’s knowledge, actions it can perform.

Abstract specification of agents

Specifying which action an agent ought to take in response to any given percept sequence provides a design for an ideal rational agent

AP*:f

Example: a thermostat

sensedtemperature

percepts action

A = {no-op,close valve, open valve}

Agent function f: T -> A, where T is set of possible ambient temperatures.

What assumptions about the environment and the deviceare we making with such an agent function?

Agent programs

An implementation of the given agent specification

function thermostat (temperature) If temperature < DESIRED – epsilon

return close valve If temperature > DESIRED + epsilon

return open valve Return no-op

AP*:f

Autonomous agents An agent is autonomous to the extent

that its behavior is determined by its own experience.

Given sufficient time and perceptual information, agent should adapt to new situations and calculate actions appropriate for those situations. Is a thermostat autonomous? Is the GPS route planner in your car

autonomous?

Taxonomy of agent programs Agents with no internal state

Reflex agents or stimulus-response agents Agents with internal state

Agents with fixed policies or reflex agents with state (agents that remember the past)

Agents that compute policies based on goals or general utility functions (agents that remember the past and can project into the future)

Agent program structure

Template for agent programs Function agent (percept) returns

action local state: l

l = update-local-state(l,percept) action = choose-best-action(l) l = update-local-state(l,action) return action

Reflex agents

Current percept

Interpret currentstate

Compute current action by choosing matching condition action rule

Action

conditionactionrules

Template for a reflex agent

Function reflexAgent(percept) static : rules, a set of condition-action

rules s =interpret-input(percept) rule = find-matching-rule(s,rules) action = rule-action(rule) return action

Does not maintain perceptual history!

An example of a reflex agent

10o

(t)

h(t)

Sensors: h(t),(t)

Performance measure: maintain heighth(t) at 3 meters (minimize the sum of (h(t)-3)2 over t in [0..T])

Actions: no-op, turninflow valve to the right, turn inflow valve to left

Condition-action rules

h 0-10 10-15

15-20

20-30

30-60

60-180

0-1.5 right right right right right noop

1.5-2.5 right right right right noop left

2.5-2.8 right right right noop left left

2.8-3.0 right noop left left left left

3.0-3.2 left left left left left left

3.2-4.0 left left left left left left

A reflex agent in nature

If small moving object then activate SNAPIf large moving object then activate AVOID and inhibit SNAP

percept

one of SNAP or AVOID

Ralph: Vision based vehicle steering sampling the image of roadway ahead

of vehicle determining the road curvature assessing the lateral offset of the

vehicle relative to the lane center commanding a steering action

computed on the basis of curvature and lane position estimates

Ralph’s sampling strategy

Curvature hypotheses

Curvature scoring

Lateral offset calculation

No hands across America

2850 mile drive from Washington DC to San Diego, on highways.

Trip challenges: driving at night, during rain storms, on poorly marked roads, through construction areas.

Evaluation metric: percent of total trip distance for which Ralph controlled the steering wheel.

Ralph steered the vehicle for 2796 of the 2850 miles (98.1 percent)

10 mile stretch of new, unpainted highway (no lane markers)

city driving when road markings were either missing or obscured by other vehicles.

Challenging highway driving

Challenging city roads

Want Ralph in your car?

Fixed video camera, forward-looking, mounted on rear-view mirror inside vehicle.

steering actuator (converts output of Ralph to steering command for vehicle).

now commercially available from Assistware Technology Inc. ($1975 from http://www.assistware.com)

Reflex agents with internal state

Current percept

Interpret currentstate

Compute current action by choosing matching condition action rule

Action

InternalState

modelof

actionsandenv

Example of reflex agent with state

An automatic lane changer: need internal state to monitor traffic in lanes unless car has cameras in the front and rear. Internal state allows one to compensate for lack of full observability.

Why internal state is useful

Or, why is remembering the past any good? So, you are not doomed to repeat it

(Santayana). Past + knowledge of actions can help

you reconstruct current state --- helps compensate for lack of, or errors in, sensory information.

Template for reflex agent with state

Function reflex-agent-with-state (percept) returns action

static: state, rules state = update-internal-state(state,percept) rule = rule-match(state,rules) action = rule-action(rule) state = update-internal-state(state,action) return action

The NRL Navigation Task

The NRL Navigation Task

A near-optimal player

A three-rule deterministic controller solves the task!

The only state information required is the last turn made.

A very coarse discretization of the state space is needed: about 1000 states!

Discovering this solution was not easy!

Rule 1: Seek GoalThere is a clear sonar in the direction of the goal.

If the sonar in the direction of the goal is clear, follow it at speed of 20, unless goal is straight ahead, then travelat speed 40.

Rule 2: Avoid Mine

Turn at zero speed to orient with the first clear sonarcounted from the middle outward. If middle sonar is clear, move forward with speed 20.

There is a clear sonar but not in the direction of the goal.

Rule 3: Find Gap

There are no clear sonars.

If the last turn was non-zero, turn again by the sameamount, else initiate a soft turn by summing the rightand left sonars and turning in the direction of thelower sum.

Optimal player in action

Where optimal player loses

Goal-based agents (agents that consider the future)

Current percept

Interpret currentstate

Compute current action by picking one that achieves goals

Action

InternalState

models

project forwardby one action from

current state

goals

Computation performed by a goal-based agent

Current state

a1 a2 a3

a4

next states projected from currentstate and action

Actions a1 and a2 lead to states that do not achieve the goal, and actions a3 and a4 do. Hence, choose one of a3 or a4.

Goal-based agents Do not have fixed policies; they

compute what to do on the fly by assessing whether the action they choose achieves the given (fixed) goals.

Are not restricted to one-step look-ahead.

Are programmed by giving them goals, models of actions, and environment.

Utility-based agents (agents that consider the future)

Current percept

Interpret currentstate

Compute current action by picking one that maximizes utility function

Action

InternalState

models

project forwardby one action from

current state

Utilityfunction

Utility-based agents vs goal-based agents

Goal-based agents are degenerate cases of utility-based agents. The utility function that goal-based agents use is:

U(s0 s1 … … sn) = 1 if sn satisfies goals = 0 otherwise

An extended example: navigation in a Manhattan grid

Case 1 Ideal sensors (robot knows where on

grid it is accurately, at all times) Ideal effectors (commanded motions are

executed perfectly) Environment: all streets two way, no

obstacles. Goal: get from (x1,y1) to (x2,y2) What kind of agent do you need to

achieve this goal?

Solution to case 1

Simple reflex agent suffices. Fixed policy: dead reckoning

Go to (x1,y2) Go to (x2,y2)

No need for sensing at all; above policy can be implemented blindly.

Case 2 Ideal sensors (robot knows where on

grid it is accurately, at all times) Real effectors (commanded motions are

not executed perfectly) Environment: all streets two way, no

obstacles. Goal: get from (x1,y1) to (x2,y2) What kind of agent do you need to

achieve this goal?

Solution to Case 2 A simple reflex agent suffices. Fixed control policy that senses position at

every time step Command motion to (x1,y2) Sense position and issue correcting motion

commands until robot is within epsilon of (x1,y2) Command motion to (x2,y2) Sense position and issue correcting motion

commands until robot is within epsilon of (x2,y2)

Case 3 Ideal sensors (robot knows where on

grid it is accurately, at all times) Ideal effectors (commanded motions are

executed perfectly) Environment: one-way streets and

blocked streets, no map. Goal: get from (x1,y1) to (x2,y2) What kind of agent do you need to

achieve this goal?

Solution to case 3 Need agent with internal state to

remember junctions and options that have already been tried there (so it doesn’t repeat past errors endlessly).

Control algorithm: shorten Manhattan distance to destination whenever possible, backing up only when at a dead end. Back up to last junction with an open choice.

Properties of environments Accessible vs inaccessible: if an agent can sense

every relevant aspect of the environment, the environment is accessible. Simple reflex agents suffice for such environments.

Deterministic vs non-deterministic: if the next state of the environment is completely determined by the current state and the action selected by the agent, the environment is deterministic. Agents with internal state are necessary for non-deterministic environments.

Discrete vs continuous: whether states and actions are continuous are discrete. Chess is discrete, taxi driving is continuous.

Properties of environments (contd.) Episodic vs non-episodic: in an episodic

environment, agent’s experience is divided into episodes. The quality of its action depends just on the episode itself --- subsequent episodes do not depend on what actions occur in previous episodes. Agents that reason about the future are unneeded in episodic environments.

Static vs dynamic: if the environment can change while the agent deliberates, the environment is dynamic for the agent. Time-bounded reasoning needed for dynamic environments.