Planning Questions and Answers

Answer: A

Demotion places the threat before the causal link and promotion places it after, and the planner branches on the two possibilities. Removing the action would usually break whatever needed it, and restoration or goal weakening are not part of the standard algorithm. Having exactly two repairs for each conflict makes the search space well defined and the algorithm straightforward to state.

Answer: B

Because only the necessary orderings are present, an executive may run independent actions concurrently or reorder them as circumstances allow, which is valuable in multi agent and resource constrained settings. Length is unchanged, monitoring remains necessary and failures still matter. A representation that records only the constraints that genuinely hold leaves the maximum freedom to whoever must act on it.

Answer: C

The arrival of relaxed plan heuristics made simple forward search outperform the more sophisticated planners on standard benchmarks by a wide margin. Partial order planning remains valuable where execution flexibility matters. Execution, expressiveness and completeness were not the issue. A more sophisticated algorithm being beaten by a simpler one with better guidance is a pattern that recurs whenever heuristics improve faster than algorithms. The quantity that varies and the quantity that does not are worth naming before any substitution is attempted.

Answer: D

An HTN planner begins with high level tasks and applies methods that decompose them into subtasks until primitive actions remain, which encodes human knowledge of how things are usually done. It scales to large practical problems at the cost of requiring that knowledge. Distributed planning, goal priorities and state abstraction are different ideas. Supplying procedural knowledge to a planner, rather than expecting it to be discovered, is what makes industrial planning feasible.

Answer: A

A method names a task, states the conditions under which the decomposition applies and gives the subtasks with their ordering constraints. Several methods for one task give the planner alternatives to search over. Action selection, costs and goal ordering are different notions. Encoding recipes rather than only physics is what distinguishes hierarchical planning from the classical formulation.

Answer: B

Writing the methods encodes how a human would tackle the domain, which is a considerable authoring effort and is why such planners solve problems that defeat classical ones. It is the same knowledge acquisition cost that expert systems faced. Plan length, completeness and verification are secondary. Trading authoring effort for solving power is a legitimate choice when the domain is stable enough to justify writing the knowledge down. Most errors here come from carrying a result across from a neighbouring case where it does not hold.

Answer: C

Fixing a horizon and encoding the initial state, the goal and the action semantics at each step gives a formula satisfiable exactly when a plan of that length exists, and the horizon is increased until one is found. The approach exploits the enormous speed of modern solvers. Solvability, precondition checking and verification are different questions. Handing a problem to a tool with better engineering, by encoding it, is a strategy worth considering whenever such a tool exists.

Answer: D

The encoding requires a horizon, since a propositional formula has finitely many variables and each time step needs its own copy of every fluent. Iterative deepening on the horizon recovers completeness. The order is what the solver determines, the initial state is given, and heuristics are internal to the solver. That the encoding is necessarily bounded, and the bound must be searched over, is the cost of the translation.

Answer: A

Each time step requires a copy of every fluent and every action variable, so a long horizon produces a formula of millions of variables. Compact encodings and parallel semantics, allowing several non-interfering actions per step, mitigate it. Negation, plan length and preconditions are handled correctly. A translation whose size grows with a parameter that must be searched over is one whose practicality depends entirely on how large that parameter must be.

Answer: B

With an uncertain initial state and no sensing, the planner must find a single sequence succeeding from every state it might be in. Such plans often exist and the search is over belief states rather than states. Safety, formats and adaptation are different notions, the last being contingent planning. That a fixed sequence can succeed under genuine ignorance is a counter-intuitive and useful result.