Planning Questions and Answers

Answer: A

Fuel is consumed and a crane is occupied and released, and a planner must respect both kinds of limit, which requires numeric reasoning rather than pure logic. Constraint and linear programming techniques are often combined with the planner for the purpose. Computing budgets, permissions and plan length are different constraints. Extending a symbolic formalism with numbers is a step that changes which solving technologies apply.

Answer: B

A planner reasons well about causal structure and poorly about arithmetic, so scheduling and resource allocation are often handed to a constraint or linear solver while the planner handles the actions. Representation is possible, plan length is a separate matter, and undecidability arises only in particular cases. Combining tools whose strengths are complementary, rather than extending one to do everything, is the same argument that produced hybrid constraint and integer solvers.

Answer: C

A planner decides what must be done and in what causal order, while a scheduler decides when each action happens and which resource performs it. Real applications need both and the boundary between them is often blurred deliberately. Timing of the computation, domain and output count are not the distinction. Separating the causal from the temporal and resource questions is what allows each to be attacked by the technology best suited to it. Framing the question as which quantity the change acts upon makes the answer follow without calculation.

Answer: D

The first two are correct. A partial order planner deliberately leaves actions unordered until a threat forces a decision, which is the principle of least commitment. Committing at every step would make it an ordinary state space planner. The three statements between them cover representation, heuristics and search strategy, which is most of what distinguishes one planner from another.

Answer: A

The four differ in what they search over: states forward, states backward, plans, and task decompositions. Naming the search space is the quickest way to predict a planner's branching factor and the heuristics that will apply to it, and it organises the whole topic more usefully than a chronological account of the algorithms would.

Answer: B

The first three are correct. A conformant plan is a single sequence that works whatever the initial state is, with no sensing at all; branching on observations is a contingent plan. The distinction determines what the executive must be capable of, and confusing the two leads to planning for a capability the agent does not have.

Answer: C

The additive, maximum and relaxed plan heuristics are all computed from the delete relaxation. Mutex propagation records interactions the relaxation deliberately ignores, so it strengthens rather than relaxes the model. Distinguishing techniques that simplify a problem to obtain a bound from those that add information to tighten one is what makes the relationship between them intelligible.

Answer: D

A plan is the action sequence, since the agent needs to know what to do rather than merely what will result. Partial order planners produce a set of actions with orderings, any linearisation of which serves. Final states, goal sets and search records are different objects. Being clear about what the planner is asked to produce determines what bookkeeping the search must carry, which is the same consideration as recovering a path in ordinary search.

Answer: A

A goal is a partial description, so any state satisfying it will do and the planner need not know which one it will reach. Requiring an exact state would be far more restrictive and rarely what is wanted. Final actions and rewards are different notions, the last belonging to decision theoretic planning. Specifying an outcome by the conditions it must meet, rather than exhaustively, is what makes goals expressible at all in a large state space. Most errors here come from carrying a result across from a neighbouring case where it does not hold.

Answer: B

Classical planning assumes the initial state is completely known, which together with the closed world assumption makes every fluent's truth determinate. Relaxing that assumption gives conformant and contingent planning. First actions and goals are different components, and partial knowledge is exactly what the classical formulation excludes. Knowing which assumptions a formalism makes is what identifies which extension is needed when they fail.