Planning
Planning is the process by which an AI agent decides what sequence of actions to take in order to achieve a goal. It differs from reactive problem solving ...
From Search to Planning
Planning is the process by which an AI agent decides what sequence of actions to take in order to achieve a goal. It differs from reactive problem solving because a planning system explicitly constructs a plan before execution rather than deciding action by action as it goes. This makes planning essential for agents operating in complex, structured environments where actions have preconditions and postconditions.
The STRIPS representation, developed at Stanford in the 1970s, is the foundational language for classical planning. Each action is described by three components: a precondition list of facts that must be true for the action to be applicable, an add list of facts the action makes true, and a delete list of facts the action makes false. For example, a pick-up action in a blocks-world domain has the precondition that a block is on the table and the arm is empty, adds the fact that the arm is holding the block, and deletes the fact that the block is on the table.
Planning algorithms search through the space of possible plans rather than the space of world states. Partial-order planning builds a plan as a set of actions with ordering constraints, adding actions and constraints only as needed to achieve the goal while avoiding conflicts. This approach is more flexible than total-order planning, which commits to a strict sequence. Modern planners like FastDownward use heuristics derived from a relaxed version of the problem to guide search, enabling them to handle problems with thousands of actions.
Hierarchical Planning
Hierarchical Task Network planning decomposes high-level tasks into progressively simpler subtasks until primitive actions are reached. A household robot might decompose "prepare breakfast" into "make coffee" and "toast bread," then decompose those into motor-level actions. This mirrors how humans naturally break down complex goals.
Interview Tip
In interviews, distinguish planning from search by noting that planning operates on a symbolic action model while generic search treats the problem as a black box. Be ready to describe a STRIPS action for a simple example, such as moving a robot through a door, by writing out its preconditions, add list, and delete list. This shows hands-on familiarity with the formalism.