FF Planning

1
FF Planning

• Fast plan generation through heuristic search.
• Enforced hill climbing for search.
• Relaxed graph plan for heuristic.
• STRIPS and ADL notation.
• Not complete.
• Does not guarantee optimal plans.

2
System Architecture
3
Relaxed Graph Plan

• Ignores delete effects of actions.
• Guarantees no exclusions.
• Causes quick graph expansion.
• Prevents backtracking.
• Allows for efficient storage.
• Stores where action first appeared.

4
Solution Length Optimization

• NOOPs first.
• Linearize by order selected.
• Difficulty heuristic.
• Sum of layers where preconditions first appear.

5
Enforced Hill Climbing

• Initialize current plan to the empty plan ltgt
• S I
• While h(S) ! 0 do
• Perform breadth first search for S' with h(S') lt
  h(S)
• If no such state is found then
• Output Fail, stop
• Add actions on path to S' to the end of current
  plan.
• S S'
• End While

6
Enforced Hill Climbing (2)

• Find the closest better successor to S.
• States are kept in a queue.
• Remove first state S' from queue and evaluate
  it with graph plan.
• If S' is better, search succeeds.
• Otherwise, add successors of S' to queue.
• If no better S' is found, search fails.

7
Completeness

• Dead End
• reachable state with no path to goal.
• Enforced Hill Climbing can solve any Dead End
  free problem.
• If Enforced Hill Climbing fails, start
  over using a complete heuristic search
  algorithm.

8
Pruning Techniques

• Helpful Actions
• Set of promising successors to search state.
• All valid actions which satisfy some goal at
  the next layer.

9
Pruning Techniques (2)

• Added goal deletion
• Prune where some goal seems achieved too
  early.
• Don't expand successors of states whose plans
  delete some goal.

10
Runtime

• Logistics World

11
Runtime (2)

• Schedule

12
Runtime (3)

• Freecell

13
Runtime (4)

• Elevator

14
Runtime (5)

• Random SAT problems

15
Individual Characteristics

• H Helpful Actions Pruning
• E Enforced Hill Climbing
• F FF heuristic estimate

16
Individual Characteristics

• H Helpful Actions Pruning
• E Enforced Hill Climbing
• F FF heuristic estimate

17
Conclusion

• FF is fast, not complete, and does not find
  optimal paths.
• Speed comes because most problems are simple.
• Completeness usually doesn't matter.
• Most downfalls would enforced hill climbing to
  fail anyways.