These refer to algorithms that operate by
solving MDP's at a series of abstraction levels, where higher levels
in the abstraction hierarchy ignore details represented at lower
levels. Hierarchical methods represent a strategy for dealing with
very large state spaces. The motivation for the use of hierarchical
models in RL is the goal of faster learning (perhaps at the expense of
slight sub-optimality in performance) through decomposition of a task
into a collection of simpler subtasks.
publications