macro-actions

A macro-action is the simplest form of temporal abstraction: a fixed, pre-set sequence of primitive actions that the agent invokes as a single unit. Think of a fighting-game combo — one button press fires off 'punch, punch, kick' in order, with no decisions in between. Choosing a macro commits the agent to running that exact list to completion before it gets to choose again.

What separates a macro from a full option is the lack of feedback while it runs. A macro is open-loop: it ignores whatever states it passes through and just replays its canned sequence, whereas an option is closed-loop, choosing each primitive action by consulting its internal policy on the current state. That makes macros simpler and cheaper to define and plan with, but brittle — if the world deviates from what the sequence assumed, the macro has no way to adapt.

Macros are useful when chunks of behaviour are genuinely repetitive and the environment is predictable enough that blind replay works: they shorten the decision horizon and can dramatically speed up search and planning. When the world is uncertain or reactive, though, the closed-loop flexibility of options earns its keep, and macros are best seen as the no-frills end of the temporal-abstraction spectrum.