Module 1: 2.1-2.4 "Intelligent Agents"

What is an agent?

Anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators.

What is a sensor?

A device through which an agent perceives its environment.

What is an actuator?

A mechanism through which an agent acts on its environment.

What is a percept?

The content an agent’s sensors are perceiving at a particular instant.

What is a percept sequence?

The complete history of everything an agent has ever perceived.

What is an agent function?

A mathematical description that maps any given percept sequence to an action.

What is an agent program?

A concrete implementation of an agent function running within a physical system.

What is the difference between an agent function and an agent program?

The agent function is an abstract mathematical mapping; the agent program is its concrete implementation.

What information can an agent's action depend on?

Its built-in knowledge and the entire percept sequence observed up to that point.

What information can an agent's action NOT depend on?

Anything it has not perceived or otherwise has no built-in knowledge about.

Why is a table-driven agent impractical?

The table can become astronomically large, making it impossible to store, construct, or learn.

What is the central challenge of AI agent design?

Produce rational behavior using a reasonably small program rather than an enormous lookup table.

What is the vacuum-cleaner world?

A simple environment where a vacuum agent perceives location and dirt and can move, suck, or do nothing.

Active recall: A vacuum agent sees that its current square is dirty. What should a simple reflex agent do?

Suck up the dirt.

Active recall: Why can two agents with identical sensors behave differently?

Their agent programs can map the same percept sequences to different actions.

What is consequentialism in the context of AI?

The idea that an agent's behavior is evaluated by the consequences it produces.

What is a performance measure?

A criterion that evaluates any given sequence of environment states according to how desirable it is.

Why is the performance measure important?

It defines what counts as successful behavior for the agent.

What determines whether a particular agent is rational?

The performance measure, prior knowledge, available actions, and percept sequence to date.

What is a rational agent?

For each percept sequence, an agent that selects an action expected to maximize its performance measure given its evidence and built-in knowledge.

What four things determine rational action?

The performance measure, prior knowledge, available actions, and percept sequence to date.

Does rationality require an agent to always achieve the best actual outcome?

No. Rationality maximizes expected performance based on available information, not actual performance in hindsight.

What is omniscience?

Knowing the actual outcome of an action and being able to act accordingly.

Why is omniscience different from rationality?

A rational agent acts using available evidence; an omniscient agent would know the actual future outcome.

Active recall: An agent makes the best choice given what it knows, but an unexpected event causes failure. Was it necessarily irrational?

No. Rationality concerns expected performance given the available information, not hindsight.

What is information gathering?

Taking actions that modify future percepts in order to obtain useful information.

Why can looking before crossing a road be rational?

The information gained can improve the expected performance of the subsequent action.

What role does learning play in rational behavior?

It allows an agent to improve or modify its knowledge based on experience.

What is autonomy?

The extent to which an agent relies on its own percepts and learning rather than the prior knowledge supplied by its designer.

When does an agent lack autonomy?

When its behavior relies heavily on the designer's prior knowledge rather than its own percepts and learning.

Why should a rational agent be autonomous?

It should learn from experience to compensate for partial or incorrect prior knowledge.

Active recall: A robot initially relies on designer rules but later learns from experience and changes its behavior. What concept does this illustrate?

Autonomy through learning.

What is a task environment?

The problem to which a rational agent is the solution; it includes the performance measure, environment, actuators, and sensors.

What does PEAS stand for?

Performance, Environment, Actuators, Sensors.

What is a PEAS description?

A specification of an agent's task environment using Performance measure, Environment, Actuators, and Sensors.

Why should PEAS be specified before designing an agent?

The nature of the task environment determines what kind of agent design is appropriate.

What does the Performance component of PEAS specify?

How successful the agent's behavior should be measured.

What does the Environment component of PEAS specify?

The external world in which the agent operates.

What does the Actuators component of PEAS specify?

The mechanisms the agent can use to act on the environment.

What does the Sensors component of PEAS specify?

The mechanisms through which the agent perceives the environment.

Active recall: For an automated taxi, name two possible actuators.

Steering and braking; other examples include accelerator, signals, horn, display, or speech.

Active recall: For an automated taxi, name two possible sensors.

Cameras and GPS; other examples include radar, speedometer, accelerometer, and engine sensors.

Active recall: Why is a taxi's performance measure likely to involve tradeoffs?

Safety, speed, comfort, legality, cost, and profit can conflict with one another.

What is a fully observable environment?

An environment where sensors give the agent access to the complete state relevant to choosing an action.

What is a partially observable environment?

An environment where sensors do not provide complete information about the state relevant to action selection.

What is an unobservable environment?

An environment in which the agent has no sensors.

Why can an environment be partially observable?

Sensors may be noisy, inaccurate, or unable to detect relevant parts of the state.

Active recall: A vacuum only senses dirt in its current square. Is the environment fully or partially observable?

Partially observable.

What is a single-agent environment?

An environment in which the agent's performance is not dependent on the behavior of other agents.

What is a multiagent environment?

An environment containing other agents whose behavior can affect the agent's performance or decisions.

Active recall: Is chess single-agent or multiagent?

Multiagent, because the opponent is another agent whose actions affect the outcome.

What is a deterministic environment?

An environment in which the next state is completely determined by the current state and the agent's action.

What is a nondeterministic environment?

An environment in which an action can have multiple possible outcomes.

What is a stochastic environment?

An environment whose possible outcomes are explicitly described with probabilities.

What is the difference between nondeterministic and stochastic?

Nondeterministic lists possible outcomes without probabilities; stochastic assigns probabilities to outcomes.

Active recall: “There is a 25% chance of rain” describes what kind of model?

A stochastic model.

Active recall: “It may rain tomorrow, but no probabilities are given” describes what kind of model?

A nondeterministic model.

What is an episodic environment?

An environment in which experience is divided into independent episodes, each involving a percept and one action.

What is a sequential environment?

An environment where current actions can affect future decisions and outcomes.

Why are episodic environments generally easier than sequential environments?

The agent does not need to consider the long-term consequences of its current action.

Active recall: Is classifying each part on an assembly line episodic or sequential if one decision does not affect the next part?

Episodic.

Active recall: Is chess episodic or sequential?

Sequential, because current moves affect future positions and decisions.

What is a static environment?

An environment that does not change while the agent is deliberating.

What is a dynamic environment?

An environment that can change while the agent is deliberating.

What is a semidynamic environment?

An environment where the physical state does not change while the agent deliberates, but the performance score can change.

Active recall: Is taxi driving static or dynamic?

Dynamic, because other cars and the world continue changing while the agent decides.

Active recall: Chess with a clock is what type of environment?

Semidynamic, because the board may remain unchanged while the clock affects performance.

Active recall: A crossword puzzle is what type of environment with respect to static/dynamic?

Static.

What is a discrete environment?

An environment with a finite or countable set of distinct states, percepts, or actions.

What is a continuous environment?

An environment involving continuously varying states, time, percepts, or actions.

Active recall: Is ordinary chess discrete or continuous?

Discrete.

Active recall: Is taxi driving discrete or continuous?

Continuous.

What does known vs. unknown describe?

The agent's or designer's knowledge of the laws governing the environment.

What is a known environment?

An environment in which the outcomes, or outcome probabilities, of actions are given.

What is an unknown environment?

An environment whose behavior the agent must learn in order to make good decisions.

Can an environment be known but partially observable?

Yes. For example, the rules of solitaire can be known even though hidden cards are not observable.

Can an environment be unknown but fully observable?

Yes. A new video game can show the complete state while the agent does not yet know what actions do.

What is the hardest combination of task-environment properties described by the textbook?

Partially observable, multiagent, nondeterministic, sequential, dynamic, continuous, and unknown.

Why does the nature of the environment affect agent design?

Different environmental properties determine what information, memory, reasoning, and decision mechanisms an agent needs.

What is a simple reflex agent?

An agent that selects actions based only on the current percept, ignoring the rest of the percept history.

What is a condition-action rule?

A rule connecting a condition derived from a percept to an action, such as “if condition, then action.”

What is the main limitation of a simple reflex agent?

It works only when the correct action can be determined from the current percept.

Why are simple reflex agents fragile in partially observable environments?

The current percept may not contain enough information to determine the true state of the world.

Active recall: A taxi sees brake lights but cannot determine whether they indicate braking or another situation. Why might a simple reflex agent fail?

It relies only on the current percept and cannot use relevant percept history.

Why can randomization sometimes help a simple reflex agent?

Random actions can help it escape infinite loops when identical percepts lead to repeated actions.

What is a model-based reflex agent?

An agent that maintains an internal state representing aspects of the world not evident in the current percept.

What is internal state?

Information maintained by an agent that depends on percept history and represents relevant unobserved aspects of the current state.

Why does a model-based reflex agent need internal state?

To keep track of aspects of the world that cannot be determined from the current percept alone.

What is a transition model?

A description of how the next state depends on the current state and the agent's action.

What is a sensor model?

A description of how the current state of the world is reflected in the agent's percepts.

How do the transition model and sensor model work together?

They allow the agent to maintain a best estimate of the current state despite limited observations.

Active recall: If a taxi knows how turning the steering wheel changes its position, what kind of knowledge is this?

A transition model.

Active recall: If a taxi knows how braking appears in camera images, what kind of knowledge is this?

A sensor model.

Can a model-based agent always know the exact state of a partially observable world?

No. It may only maintain a best guess or multiple possible states.

What is a goal-based agent?

An agent that uses goal information together with a model to choose actions that achieve its goals.

What is a goal?

Information describing desirable situations the agent is trying to achieve.

Why is a goal-based agent more flexible than a simple reflex agent?

Its explicit goal information can be changed without replacing all of its condition-action rules.

How does a goal-based agent consider the future?

It predicts what will happen after possible actions and chooses actions that eventually achieve its goals.

Why are search and planning relevant to goal-based agents?

They find sequences of actions that can achieve goals when a solution is not immediate.

Active recall: A taxi can turn left, right, or straight, and the correct choice depends on its destination. Which agent type handles this naturally?

A goal-based agent.