Planning Questions and Answers

Answer: A

Preconditions gate applicability, so a planner may apply an action only in states satisfying them, and achieving them becomes a subgoal in backward search. Effects describe what becomes true afterwards. Hopes and goal conditions are different notions. Separating what must hold before from what holds after is the whole structure of an action description, and every planning algorithm is organised around that division.

Answer: B

The add list names what becomes true and the delete list what becomes false, and everything unmentioned persists. The delete relaxation ignores the second list, which is what makes the relaxed problem monotonic. Preconditions and unchanged fluents are different categories. Knowing which list a heuristic ignores, and why that makes the relaxed problem easy, connects the representation directly to the heuristics computed from it.

Answer: C

The delete list names the fluents that cease to hold, and ignoring it gives the relaxation from which almost every modern planning heuristic is derived. Removing a fluent from the domain, negative preconditions and unknown values are different notions. That one list of a representation can be dropped to obtain a tractable relaxation is a happy accident of the formalism, and it is worth appreciating how much rests on it.

Answer: D

Monotonic growth means an action applied now can never prevent one applied later, so a greedy fixed point computation finds every achievable fluent without search. Action counts, goal satisfaction and preconditions are unchanged. Recognising monotonicity as the property that removes the need for search, here as in forward chaining over definite clauses, is what makes the relaxation's tractability intelligible rather than surprising.

Answer: A

The relaxation heuristics turned greedy best first forward search into the dominant approach on standard benchmarks, displacing far more sophisticated algorithms. Complexity is unchanged, partial order planning remains correct and useful, and domain descriptions are still required. That better guidance rather than a better algorithm produced the advance is a lesson that applies to search generally and is worth carrying beyond this topic.

Answer: B

The standard architecture grounds the problem, computes a relaxation based heuristic at each state and searches forward greedily, with enhancements for plateaux and dead ends. Satisfiability encodings and partial order planners remain in use for particular purposes. Exhaustive enumeration is hopeless. Knowing the default architecture, and the circumstances under which an alternative is preferred, is what allows a planner to be selected rather than assumed.

Answer: C

On a region where every successor has the same estimate, greedy search has no guidance, so an exhaustive local search is used until a strictly better state is found. Enforced hill climbing is the standard technique. Abandoning, restarting and accepting are all incorrect responses. Providing a defined fallback for the situation in which the primary mechanism gives no information is what distinguishes a robust implementation from a demonstration.

Answer: D

In domains with irreversible actions a state may be reached from which no plan exists, and a planner unable to detect that will search beneath it indefinitely. Detecting dead ends cheaply is an active research problem. Repeated states, inapplicability and thresholds are different conditions. A failure mode that arises only when actions cannot be undone is worth knowing about when choosing a planner for a domain in which they cannot.

Answer: A

In a reversible domain any mistake can be undone, so greedy search recovers, whereas an irreversible action may commit the agent to a state with no plan. Breaking a fragile object is the standard example. Representation and heuristics are unaffected, and irreversibility does not in itself create a dead end. Asking whether mistakes are recoverable, before choosing a search strategy, determines how cautious that strategy must be.

Answer: B

Where some actions cause permanent harm, an agent must recognise and avoid them while still exploring enough to learn, which is a genuinely difficult requirement. It matters most in physical systems where a mistake damages equipment or people. Visited states, bounded regions and spare capacity are different restrictions. A requirement arising from the consequences of action rather than from the algorithm is one that the algorithm must be extended to respect.