Intelligent Agents Questions and Answers

Answer: A

A reactive agent maps perception to action with little or no intervening computation, which gives speed and robustness at the cost of foresight. Subsumption architectures are the canonical example. Change driven operation, social response and replanning are different behaviours sharing the word. Reactive and deliberative are the two poles that hybrid architectures try to combine, and being clear about which is meant keeps that discussion coherent.

Answer: B

A deliberative agent maintains a model and searches or plans over possible futures, which lets it pursue extended goals but makes it slow. Slowness is a consequence rather than the definition, joint deliberation is a multi-agent activity, and authorisation is a constraint. The contrast with reactive design is between anticipating consequences and responding to conditions, and a serious system usually needs both operating on different timescales.

Answer: C

Every agent runs a loop in which it senses, updates whatever internal state it keeps, chooses an action and executes it, and the rate of that loop is one of the most consequential design parameters. Latency, polling order and update frequency are details within it. Recognising that all agent architectures share this loop and differ only in what happens inside it makes the succession of designs much easier to compare.

Answer: D

In a dynamic environment a decision based on a stale percept may be acted on in a state that no longer exists, so the cycle must keep pace with the rate at which the world changes. A faster cycle in a static environment buys nothing and costs computation. Memory and regulation are separate matters. Matching a system's response rate to the timescale of its environment is the defining requirement of real time design, and it is set by the task rather than by preference.

Answer: A

Fusion combines complementary and redundant measurements, weighting each by its reliability, to give a better estimate than any one sensor could provide, and a Kalman filter is the classical implementation. Physical connection, sensor replacement and switching are different arrangements, though selection is a crude form of fusion. Using the disagreement between sensors to estimate their reliability, as well as the state, is what makes a fusion algorithm more than an averager.

Answer: B

With noisy or ambiguous percepts the agent cannot know its state, so it maintains a probability distribution over the possibilities and updates it as evidence arrives. Acting on the most likely state alone discards information that matters when the alternatives have very different consequences. Storage, actuator behaviour and rate of change are separate concerns. Representing what is not known, rather than guessing and proceeding, is what distinguishes a probabilistic agent from a deterministic one.

Answer: C

A belief state assigns a probability to each state the agent might be in, and is updated by Bayes' rule as percepts arrive and by the transition model as actions are taken. In a logical setting the corresponding notion is a set of possible worlds. Nested beliefs about other agents and deterministic internal state are related but distinct. Carrying a distribution rather than a point estimate is what allows the agent to act sensibly when it is genuinely unsure.

Answer: D

Acting moves probability mass according to how the action changes the world, generally spreading the distribution because the outcome is uncertain, and the subsequent percept then sharpens it again. Prediction and correction alternate in exactly this way in every filtering algorithm. Resetting, collapsing and freezing the distribution all discard information. The alternation of a step that increases uncertainty with one that reduces it is the structure of every state estimator.

Answer: A

The transition model supplies the prediction step, saying how likely each next state is given the current state and the action, while the sensor model supplies the correction step. Action selection and performance measurement are separate functions. Naming the two models and the two steps precisely makes every filtering algorithm, from the Kalman filter to a particle filter, recognisable as the same structure with different representations.

Answer: B

The sensor or observation model gives the likelihood of seeing a particular percept if the world is in a particular state, which is exactly what Bayes' rule needs to convert an observation into a belief update. Physical specification informs it but is not it, evolution is the transition model, and attention is a separate control problem. Writing the model as a conditional probability in the right direction is the step that most often goes wrong when it is first attempted.