Intelligent Agents Questions and Answers

Answer: C

The performance measure lives in the user's head and is never stated, so the agent must infer preferences from sparse and ambiguous evidence and will be judged against a standard it cannot see. Speed, storage and integration are engineering problems with known solutions. A task whose objective must itself be learned is fundamentally harder than one where it is given, and it connects directly to the alignment problem in its practical form.

Answer: D

The performance measure is the external standard by which a designer or user judges the agent's behaviour, while the utility function is the internal quantity the agent maximises, and the two coincide only if the design is right. The gap between them is exactly where specification gaming lives. Task type and authorship are not the distinction. Keeping the external criterion and the internal objective conceptually separate is what allows the difference between them to be discussed at all.

Answer: A

A capable optimiser pursues the utility it holds with complete fidelity, so any divergence from what was actually wanted is pursued just as vigorously, and the more capable the agent the more thoroughly it exploits the divergence. Computation cost, documentation and learning are separate issues. That capability amplifies the consequences of a misspecified objective, rather than compensating for it, is the reason the distinction matters increasingly.

Answer: B

Internal state is a compression of the percept history that retains what matters for future decisions and discards the rest, which is what makes a model based agent feasible where a table driven one is not. Storing everything is unnecessary, the unobserved environment is by definition not directly available, and goals are held separately. Deciding what to remember is itself a design question, and getting it wrong shows up as an agent that cannot distinguish situations it needs to.

Answer: C

Updating internal state needs a transition model saying how the world changes on its own and how it changes in response to the agent's actions. Together they let the agent maintain a picture of what it cannot currently see. Sensor and actuator models refine the same idea, goals and utilities belong to higher architectures, and other agents are a special case of world evolution. Naming exactly what knowledge an architecture requires shows immediately how much work building one will be.

Answer: D

Given a model and a goal, the agent searches or plans over sequences of actions to find one reaching the goal, which makes it flexible when the goal is changed. A utility function is the next step up and handles conflicting or graded objectives. Learning and reflex rules belong to other architectures. Separating the objective from the world knowledge, so that either can be changed without the other, is the structural advance this architecture makes.

Answer: A

A binary goal cannot express that arriving quickly is worth some risk, or that one acceptable outcome is much better than another, so a utility function is needed. Large state spaces, time pressure and sensor noise are handled within either architecture by other means. Recognising that the inadequacy lies in what can be expressed, rather than in what can be computed, identifies which addition is actually required.

Answer: B

With probabilities from a model and utilities over outcomes, the agent can compute the expected utility of each action and choose the highest, which is the standard formulation of rational choice under uncertainty. Sensor reliability, search efficiency and guarantees come from elsewhere. Combining a probability distribution with a preference ordering into a single comparable number is what makes decision theory operational, and it is the formal core of the rational agent view.

Answer: C

The principle weighs each possible outcome by its probability and its utility and selects the action whose average is greatest, which follows from a small set of axioms about consistent preferences. Considering only the most likely outcome ignores rare catastrophes, avoiding bad worst cases is the more conservative minimax criterion, and ignoring probability abandons the uncertainty altogether. A decision rule derivable from axioms about preference is on firmer ground than one adopted by intuition.

Answer: D

The number of outcomes is usually far too large to enumerate, and the probabilities and utilities are themselves estimates, so approximations, sampling and satisficing are used instead. Expected utility applies to sequential problems too, probabilities can often be estimated, and risk attitude is captured within the utility function itself. A normative principle that cannot be computed exactly is still valuable as a target that approximations are measured against.