Intelligent Agents Questions and Answers

Answer: A

A concave utility means that each additional unit of money or reward is worth less than the last, so a certain amount is preferred to a gamble of the same expected value. That is how risk attitude enters without abandoning expected utility maximisation. Distorting probabilities, adding penalties and refusing uncertainty are ad hoc alternatives that lack the axiomatic justification. Absorbing an apparent departure from the theory into the shape of the utility function keeps the framework intact.

Answer: B

The critic observes the percepts and reports performance against a fixed external standard, since the agent's own utility function cannot be used to judge whether that utility function is right. Without an external standard the agent could improve its score by changing its opinion of what counts. Architecture, algorithm selection and exploration belong elsewhere. Requiring the feedback to come from outside the system being improved is what makes the learning meaningful.

Answer: C

If the agent could modify the criterion by which it is judged, improvement would become trivial and meaningless, so the standard must be fixed and outside its control. This is the same reason a performance measure is set by the designer rather than by the agent. Accuracy, access and regulation are secondary. A feedback loop in which the system can alter its own target is not a learning loop at all, which is a point that applies well beyond agent design.

Answer: D

The first two are correct. Handling conflicting goals is precisely what a utility function is for, since it converts incomparable objectives into a single ranking. A goal based agent is the one that cannot resolve conflicts, having only a binary criterion. The three statements together trace the ladder of architectures, and the third inverts the very capability that motivates the last rung.

Answer: A

Each dimension of the environment taxonomy selects a different capability, which is exactly why the classification is worth doing before any design work begins. Memory, probability, timeliness and strategy are four distinct requirements, and a task at the difficult end of every dimension needs all four. Driving is the standard example of exactly that, which is why it has proved so much harder than the early demonstrations suggested.

Answer: B

The first three are correct. Constraint propagation requires variables to propagate between, so it needs a factored representation at least, and an atomic representation supports only equality tests. The hierarchy matters because each level admits a class of algorithm the level below cannot express, so choosing a representation commits the design to a set of available techniques before any of them is selected.

Answer: C

Observability, determinism and episodicity are properties of the task environment and are settled by the problem rather than the designer. Symbolic against connectionist is a choice of method, made by the designer in response to the environment. Keeping the properties of the problem separate from the choices about the solution is the discipline the classification exists to enforce, and confusing the two leads to designs justified by preference rather than by the task.

Answer: D

Both are true and the second is the reason for the first. If two states look identical to the sensors but demand different responses, no function of the current percept can be correct in both, so a memory of earlier percepts is logically necessary. It is an argument from impossibility rather than from efficiency, which makes it much stronger. Establishing that an architecture cannot work, rather than that it works badly, settles the design question immediately.

Answer: A

The first is correct: where the percept reveals the whole state, a condition action rule can always be correct. The second fails because the same percept may arise in states requiring different actions, and no rule keyed on the percept alone can be right in both. The pair shows that architecture adequacy is a property of the architecture and the environment together rather than of the architecture alone, which is the central lesson of the whole classification.

Answer: B

The action space is what the agent can do, and its size and structure determine how hard action selection is: a discrete space of a few options admits exhaustive comparison, while a continuous space requires optimisation. The reachable states are a consequence of the actions rather than the actions themselves. Space and time are separate constraints. Characterising both the percept and the action spaces is part of specifying a task before any design begins.