jonathon doran. the planning domain a domain describes the objects, facts, and actions in the...

27
Unifying Planning Techniques Jonathon Doran

Upload: brandon-benson

Post on 12-Jan-2016

218 views

Category:

Documents


4 download

TRANSCRIPT

Page 1: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Unifying Planning Techniques

Jonathon Doran

Page 2: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

The Planning DomainA domain describes the objects, facts, andactions in the universe.

We may have a box and a tablein our universe.

Page 3: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Fluents

Propositions that aretrue or false at a pointin time.

There is a green box.The box is on the table.

The state of the universe is describedby a set of these fluents.

Page 4: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Actions

A set of preconditionsA set of fluents to addA set of fluents to delete

STRIPS formalism

Our goal is to move the universefrom one state to another.

Actions change the universe’s state.

Page 5: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Plans

The goal is typically a subset of a final state.

Actions imply state... plans are equivalently a list of states.

An initial state of the universe, and a list of actionsthat change the universe to some goal state.

Page 6: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Plan Structure

We could look at a plan as an alternatingsequence of states and actions...

SASASAS ...

Plans are also recursive.

SA (S A (S A S))

Plans are not necessarilyunique.

Page 7: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

The State Space

Actions:

Up, Down, Left, Right

This space growsexponentially withthe number of variables!

Page 8: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Complications

How do we plan a route?

We cannot reliably estimate our distance to the goal...

Page 9: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Planning is NP-Complete

Abandon all hope

It gets worse:

Optimal planning is NP-Hard

Early planners did an exhaustive search.

Page 10: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Some HopeBut we have seen techniques fordealing with NP-complete problems...

DPLL prunes the solution space. This works if we are

lucky.

Page 11: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

GraphPlan

Blum and Furst1995

Represents solution spaceas a graph.

Able to prune this graphin polynomial time.

Able to build this graph in polynomial time.

Page 12: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Graph Representation

Circles represent fluents.Squares represent actions.

Page 13: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Mutual Exclusion

Mutually exclusive situations maybe identified and pruned.

Actions are mutex if no plan couldcontain both at the same time.

Page 14: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Mutex ExamplesInterference: (one action deletes a precond of another)

I have $15, and can either buy a CD or a movie ticket.

Inconsistent Effects:

The effect of one action is the negation of another action.

I wish to paint the room entirely blue and entirely red.

Page 15: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Mutex ExamplesCompeting Needs:

Two actions have preconditions that are mutex.

I wish to sleep and go to class.

Page 16: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Fluent MutexesFluents are mutex if all ways of asserting them are mutex.

I am presently in Dallas and Denton.

What would it mean for some actions(but not all) to be mutex?

Or, if one fluent is the negation of another.

Page 17: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Graph GrowthActions monotonically increase.

Fluents monotonically increase.

Fluent-mutexes monotonically decrease.

Action-mutexes monotonically decrease.

At some point the graph stops changing.

Page 18: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Theorem

The size of the graph, and the time to create it are:

• polynomial in t (the number of levels)• polynomial in n (the number of objects)• polynomial in m (the number of actions)• polynomial in p (the number of fluents in the goal)

O(k(a+n)^2)

Page 19: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

SatPlanExpress the problem as a CSP (Boolean Sat for example)

UseAction(a1, t1) ^ UseAction(a2,t2)...

Holds(p1,t1) ^ Holds(p2,t2) ...

With appropriate constraints.

Run this through a Sat Solver

Page 20: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Unifying GraphPlan and SatPlanCreate a planning graph, like GraphPlanfor k-levels.

Convert graph to a CNF wff

Run through a Sat Solver.

Increase graph to k+1 levels if needed

Page 21: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

A Good Idea?A wff from a graph is much smallerthan using STRIPS operators.

Random restarts used in implementations.

Restart threshold gradually increases.

SatPlan requires significantpreparation time.

Page 22: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Results

Page 23: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

Rocket Example Problemtwo rockets: X1, X2locations: Earth, Mars, Venuspeople: Adam, Bettyboth rockets fueled, and at EarthAdam and Betty on Earth

Goal: Adam at Mars, Betty at Venus

Page 24: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

DomainPredicates: (at <X> <planet>) (in <X> <rocket>)

(fueled <rocket>)

Actions: (load <person> <rocket>)(unload <rocket> <planet>)(fly <rocket> <planet>)

Page 25: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

16 Level 1 Actions(load A 1) (load A 2) (load B 1) (load B 2)(F 1 E) (F 1 M) (F 1 V) (F 2 E) (F 2 M) (F 2 V)

Plus 6 no-op actions

Page 26: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

14 Level 1 Fluents(at A E) (in A 1) (in A 2)(at B E) (in A 1) (in A 2)(at X1 E) (at X1 M) (at X1 V)(at X2 E) (at X2 M) (at X2 V)(fueled 1) (fueled 2)

Page 27: Jonathon Doran. The Planning Domain A domain describes the objects, facts, and actions in the universe. We may have a box and a table in our universe

ResearchDirectionsplanning with preferencescooperative planningdistributed planningprobabilistic planningplanning under uncertainty (SCM, POMDP)