planning

In depth

Two regimes

Planning consists in turning a vague goal into executable steps, with an order and dependencies. Two regimes coexist. In the first, the plan is written in one block at the outset then followed: the trajectory is legible and can be budgeted, but the plan ages from the first surprise. In the second, the system plans only one step at a time and decides what comes next in the light of what it has just observed, which makes it more robust, less predictable and harder to cost. Most systems mix the two, with an initial plan revised as the turns go by.

An inspectable plan

An explicit plan has a virtue that is underestimated: it makes the reasoning inspectable before anything at all is executed. It can be read, corrected, refused, or put to a person for approval when the planned actions are irreversible. It also opens an axis of measurement, since comparing the announced plan with the path actually taken reveals drift far better than the final result alone. Finally, it makes it possible to set off again from an intermediate point rather than relaunching everything from the beginning.

The traps

The first trap is to mistake a well written plan for a feasible one: a model readily produces elegant steps that presuppose tools which do not exist or information it does not hold. The second is the plan that is never revised, followed to the end although its first step has already failed. The third is excessive breakdown, where each step costs an extra call and where the sum of the calls far exceeds the gain obtained. One rule holds: a plan serves to decide what to do next, not to reassure the person reading it.