Intelligent Agents Questions and Answers

Answer: A

The number of possible percept histories grows exponentially with the length of the agent's life, so even a chess playing table would need more entries than there are atoms in the observable universe. Lookup itself is fast, which is the only attraction of the design. Tables represent percept dependent behaviour perfectly well, and goal change is a separate objection. A design that is correct in principle and impossible in practice is still worth stating, because it defines what every real architecture is approximating.

Answer: B

The whole of agent design is finding compact procedures that reproduce the behaviour a vast table would have specified, and every architecture is a different way of doing so. Learning, internal state and lookahead are particular techniques rather than the general distinction, and simple reflex agents use none of them. Recognising the entire field of agent architecture as an exercise in compression makes the succession of designs easier to follow.

Answer: C

In an atomic representation states can only be compared for equality, so search algorithms can visit and remember them but cannot reason about what they have in common. It suffices for basic search and for Markov models. Factored representations use variables and structured ones use objects and relations. The choice of representation determines which algorithms are available, which is why it is settled before any algorithm is chosen.

Answer: D

Splitting a state into variables lets an algorithm notice that two states agree on most variables, which is what makes constraint propagation, Bayesian networks and planning heuristics possible. Objects and relations require the further step to structured representations. Memory efficiency is a side benefit and dynamism is unrelated. Each level of representation buys a class of inference that the level below cannot express, which is the reason the hierarchy is worth learning.

Answer: A

Expressing a relation between two objects requires a structured representation with objects, relations and the ability to quantify over them, which is what first order logic and relational planning languages provide. A factored representation can encode a fixed set of blocks clumsily but cannot generalise over them. Atomic representations cannot express it at all. The choice is not merely convenience: it determines what can be said as well as what can be computed.

Answer: B

A percept is the raw input at one moment, and the percept sequence is the whole history of them. Interpretation, action selection and internal state are things the agent does with percepts rather than percepts themselves. Being precise about the vocabulary matters because the agent function is defined as a mapping from percept sequences to actions, and a loose reading of either term makes that definition useless.

Answer: C

With two squares and two conditions, the whole environment has four states and the agent function can be written out as a table, so every architecture can be compared exactly. Real vacuum robots are far more complex, the performance measure is notoriously easy to get wrong, and internal state is needed as soon as the environment is made partially observable. A worked example small enough to enumerate is worth a great deal when the general case is intractable.

Answer: D

Rewarding the quantity collected rewards a cycle of dumping and recollecting, whereas rewarding a clean floor over time rewards the outcome actually wanted. It is the classic illustration of specifying the measure in terms of what is desired rather than what is easy to count. The other options describe behaviours that this measure does not encourage. An objective that admits a degenerate solution will find it, reliably, whenever the optimiser is any good.

Answer: A

Chess with a clock is semidynamic: the board waits for the player but the performance measure penalises delay, so time matters even though the state does not change. A fully dynamic environment changes independently of the agent. The other options describe static, quasi-static and partially dynamic notions that are not the standard term. A category defined by whether the score rather than the state depends on time is a fine distinction that turns out to matter for real time design.

Answer: B

In an episodic task the choice in one episode does not affect later ones, as in classifying items on a conveyor, so no lookahead is needed. In a sequential task, as in chess or driving, present choices have long term consequences and planning becomes essential. Resetting, discrete perception and end of run scoring are different notions. Whether decisions interact is one of the strongest determinants of how hard a task is.