Hierarchical RL & Options

temporal abstraction

Temporal abstraction is the idea of reasoning about chunks of time rather than single ticks of the clock. When you plan a trip you think 'fly to Tokyo, take the train, walk to the hotel' — three decisions covering a whole day, not the millions of muscle movements inside them. Letting an agent decide in terms of extended behaviours, each spanning many primitive steps, is exactly this kind of coarsening of the time axis.

It is the unifying principle behind every method in this field: options, skills, subgoals, and manager–worker hierarchies are all concrete ways to build temporally abstract actions. The benefit is mechanical and large. A decision that commits to a behaviour for k steps shrinks the number of choices the high level must make by roughly a factor of k, and it shortens credit assignment, because reward must only be traced back across a handful of abstract decisions rather than across the full primitive horizon.

The catch is choosing the right grain. Too coarse and the abstract actions cannot express what the task needs; too fine and you regain none of the benefit. Good temporal abstractions also have to align with the structure of the world — they work because real tasks genuinely decompose into reusable, repeatable pieces.

Also called
action abstraction over time