Planning Questions and Answers

Answer: A

Causal reasoning suits a planner and numeric temporal reasoning suits a constraint or linear solver, so a plan is generated and then scheduled, with feedback if the schedule is infeasible. Solving jointly is possible and sometimes better but far harder. Correctness, timing and domains are not the reason. Decomposing a problem along the boundary between two mature technologies, and iterating between them, is a pragmatic architecture that appears throughout operations research.

Answer: B

Feedback from the scheduler to the planner, requesting a plan avoiding the bottleneck, is the standard arrangement, and the loop repeats until a feasible pair is found. Accepting violations and relaxing limits change the problem, and declaring insolubility is premature. An architecture in which two components negotiate, rather than one simply consuming the other's output, is what makes decomposition workable when the parts are not truly independent.

Answer: C

Handing the causal part to a planner and the numeric part to a scheduler lets each use decades of specialised development, at the cost of iterating when the decomposition is imperfect. Complexity is unchanged and optimality is generally lost. Organisational separation is a side effect rather than a benefit. Accepting a loss of optimality in exchange for being able to use two mature tools is a trade that should be made explicitly rather than by default.

Answer: B

The statement that is not a property is Actions decrease monotonically.

In a planning graph (as in GraphPlan), literals increase monotonically: once a literal appears at a level, it persists through no-op actions to all later levels. Actions increase monotonically too: once an action's preconditions appear, they remain, so the action stays available. Mutexes decrease monotonically, because if two items are mutually exclusive at a level they were also mutex at every earlier level.

Since literals and actions only grow and mutexes only shrink, the graph eventually levels off when two consecutive levels are identical. So actions do not decrease.

Answer: C

The statement not true of STRIPS is It uses open world assumption.

STRIPS uses the closed world assumption: any condition not mentioned in a state is taken to be false. It is ADL (Action Description Language) that adopts the open world assumption.

The other statements are true of STRIPS: states contain only positive literals, goals and effects are conjunctions, and there is no built-in support for equality or types, which ADL adds.

Answer: A

Sensorless planning is also known as Conformant planning.

In sensorless (conformant) planning the agent cannot observe the world, so it must find a fixed sequence of actions that reaches the goal whatever the initial state or the outcomes of its actions turn out to be. It reasons over belief states, the sets of states it might be in.

Conditional or contingency planning is the opposite case: the agent can sense, and the plan branches on what it observes. Continuous planning keeps revising the plan as the world changes while it executes.

Answer: B

Forward state space search is also known as progression planning.

A forward planner starts from the initial state and applies actions whose preconditions are satisfied, progressing through successor states until it reaches a state that satisfies the goal.

Regression planning is the opposite, backward search: it starts from the goal and regresses through actions to subgoals until it reaches the initial state. Contingency planning prepares plans with branches for different observations in uncertain environments, and 'consistency planning' is not a standard term.

Answer: C

The kind of backward chaining in which operators are selected and sub goals are set up to establish the preconditions of the operators is called operator subgoaling. In means-end analysis, once a difference is found, an operator that reduces it is chosen, and if it cannot yet be applied, the aim of making its preconditions true becomes a sub problem.

Goal stack planning and backward planning are different planning methods, and operator overloading is a programming language feature.

Answer: B

The blocks world problem is normally discussed to explain a planning system. In it a robot arm has to rearrange blocks on a table using operators such as pick up, put down and stack, and the aim is to find a sequence of actions from the start to the goal arrangement, which is a planning task.

It is solved with search, but the problem is used as the standard planning example, not to teach search techniques as such. It is not a constraint satisfaction problem and it is not a question of building a knowledge base, so the other options are wrong.

Answer: D

In means-ends analysis, once a difference between the current state and the goal state is found, one applies an operator that can reduce the difference. If the operator cannot be applied at once, a subgoal is set to meet its conditions, and the process repeats.

Forward, backward and bidirectional search describe the direction in which a search is carried out. Means-ends analysis chooses operators on the basis of the difference, so these options do not describe the step.