Hierarchical Task Networks

• Each task can be accomplished by one or more "methods".
• Each method, in turn, specifies lower-level tasks to be accomplished.
• At the bottom-level tasks, actions are added to the plan.
• Encodes a domain-specific planning algorithm.

Example HTN Planner: Pyhop

• Pyhop is written as a Python library
• Operators and methods are simply Python functions
• States are Python data structures
• Pyhop algorithm:
  • Initialize task list to contain the top-level task in the hierarchy.
  • Initialize plan to empty
  • Seek-Plan(state, tasks, plan)
    • If tasks is empty, return plan
    • Examine the first task from tasks
      • If this task is an operator
        • Apply the operator to update the state
        • Call Seek-Plan(newstate, the remaining tasks, plan + operator) and return the resulting plan (if found)
      • If this task is a method
        • For each relevant method
          • Call Seek-Plan(state, subtasks + remaining tasks, plan) and return the resulting plan (if found)
    • Return with failure
• Observations
  • Performs depth-first search
  • Does not attempt to find optimal plans
  • Only backtracks due to failure