Classical Planning and GraphPlan

1. Classical Planning and GraphPlan Classes 17 and 18 All slides created by Dr. Adam P. Anthony

2. Overview • Planning as a separate problem • Planning formalism • Example planning problems • Planning in State Space • Planning Graphs • GraphPlan Algorithm • Other approaches

3. Planning as a Separate Problem • Planning = Determining a sequence of actions to take that will achieve some goal(s) • Doesn’t that sound familiar? • That sounds similar to a search algorithm • Also, we could re-cast this idea using first-order logic • Key insight: planning domains are much more carefully structured and constrained than the general search/resolution problems • A customized algorithm will have better performance

4. Planning Domain Definition Language • PDDL is a factored representation • Entire world represented by variables • In essence PDDL defines a search problem: • Initial state • Actions (with preconditions) • Action results (post conditions) • Goal test/State • Similar to traditional language called STRIPS

5. Representing States • Use a language similar to FOL: • Poor ^ Unknown • At(Truck1, Melbourne) ^ At(Truck2, Sydney) • Difference #1 from FOL: Database semantics • Anything not mentioned presumed false (no negations needed/allowed) • Unique names certain to specify distinct objects • Difference #2: all facts are ground, and functionless • At(x,y) not permitted, nor is At(Father(Fred), Sydney) • Can be specified using logic notation (common) or with set semantics where fluents are categorized into groups that are manipulated using set operators Each fact is called a Fluent

6. Representing Actions • Actions are specified in terms of what changes • Anything not mentioned is presumed to stay the same • Uses an action schema • Limited FOL representation (all universally quantified) • Action name • Variable list • Precondition • Effect • Taking an action constitutes successfully grounding all the variables to literals that are true (or not) in the state At(here) ,Path(here,there) Go(there) At(there) , At(here)

7. Sample Action Action Schema: Action(Fly(p,from,to), PRECOND: At(p,from)  Plane(p)  Airport(from)  Airport(to) Effect: At(p,from)  At(p,to)) Actual (Grounded) Action: Action(Fly(P1,SFO,JFK), PRECOND: At(P1,SFO)  Plane(P1)  Airport(SFO)  Airport(JFK) Effect: At(P1,SFO)  At(P1,JFK)) Number of grounded Actions: O(Vk) where V is the number of variables in the action, k is the number of literals defined in the state

8. Applying Actions • An action is applicable if the preconditions are satisfied by some literals in the state S • Applying an action a in state S, has the result S’: • S’ = (S – Del(a))  Add(a) • Add(a): Add-list comprised of positive literals in a’s effects list • Del(a): Delete-list comprised of negative literals in a’seffects list • To remain consistent, we require that any variable in the effects list also appear in the preconditions list • Time is implicit in the language: Actions are taken at time T, and effects occur at time T+1

9. Planning Domains • A set of action schemas completely specifies a Planning Domain • A single Planning Problem includes all the schemas from the domain, plus an initial state and a goal • Initial state: any conjunction of ground atoms such that each atom either appears, or can be bound to a variable, in at least one precondition item for at least one action • Goal: Many ways of veiwing. Simplest: Action schema where goal test is the precondition and the effect is the ground literal: GoalAccomplished • Multiple Goals are covered by making one more goal in which accomplishing each goal is part of the precondition

10. Example Planning Domains • Cargo transport • Spare tire problem • Blocks World

11. Example: Cargo Transport • Reality: Fed-Ex and UPS • Simplification: • Actions = Load, Unload, Fly • Predicate: In(c,p): package c is in plane p • Predicate: At(x,a): Item (plane or package) is At airport a • Packages are no longer ‘at’ the airport, if they are ‘in’ the plane (to compensate for lack of universal quantifiers) • In Class: define the action schema

12. Example: Spare Tire Problem • Goal: restore a car to having 4 good tires • Fluents: Tire(Flat), Tire(Spare), Trunk, Axle, Ground • Predicate: At(x,y) • Actions: • Remove(obj, loc) • PutOn(obj,loc) • LeaveOvernight—all tires are stolen • In Class: define actions and discuss whether domain is realistic enough to be useful

13. B A C TABLE Example: Blocks world The blocks world is a micro-world that consists of a table, a set of blocks and a robot hand. Some domain constraints: • Only one block can be on another block • Any number of blocks can be on the table • The hand can only hold one block Typical representation: ontable(a) ontable(c) on(b,a) handempty clear(b) clear(c)

14. Blocks world operators • Here are the classic basic operations for the blocks world: • stack(X,Y): put block X on block Y • unstack(X,Y): remove block X from block Y • pickup(X): pickup block X • putdown(X): put block X on the table • Each will be represented by • a list of preconditions • optionally, a set of (simple) variable constraints • The effects, split into ADD and DEL: • a list of new facts to be added (add-effects) • a list of facts to be removed (delete-effects) • For example: preconditions(stack(X,Y), [holding(X), clear(Y)]) deletes(stack(X,Y), [holding(X), clear(Y)]). adds(stack(X,Y), [handempty, on(X,Y), clear(X)]) constraints(stack(X,Y), [XY, Ytable, Xtable])

15. Blocks world operators II operator(stack(X,Y), Precond [holding(X), clear(Y)], Constr [XY, Ytable, Xtable], Add [handempty, on(X,Y), clear(X)], Delete [holding(X), clear(Y)]). operator(pickup(X), [ontable(X), clear(X), handempty], [Xtable], [holding(X)], [ontable(X), clear(X), handempty]). operator(unstack(X,Y), [on(X,Y), clear(X), handempty], [XY, Ytable, Xtable], [holding(X), clear(Y)], [handempty, clear(X), on(X,Y)]). operator(putdown(X), [holding(X)], [Xtable], [ontable(X), handempty, clear(X)], [holding(X)]).

16. STRIPS planning • STRIPS: first major planning system out of SRI • STRIPS maintains two additional data structures: • State List - all currently true predicates. • Goal Stack - a push down stack of goals to be solved, with current goal on top of stack. • If current goal is not satisfied by present state, examine add lists of operators, and push operator and preconditions list on stack. (Subgoals) • When a current goal is satisfied, POP it from stack. • When an operator is on top stack, record the application of that operator on the plan sequence and use the operator’s add and delete lists to update the current state.

17. A C B A B C Typical BW planning problem Initial state: clear(a) clear(b) clear(c) ontable(a) ontable(b) ontable(c) handempty Goal: on(b,c) on(a,b) ontable(c) • A plan: • pickup(b) • stack(b,c) • pickup(a) • stack(a,b)

18. A C B A B C Another BW planning problem Initial state: clear(a) clear(b) clear(c) ontable(a) ontable(b) ontable(c) handempty Goal: on(a,b) on(b,c) ontable(c) • A plan: • pickup(a) • stack(a,b) • unstack(a,b) • putdown(a) • pickup(b) • stack(b,c) • pickup(a) • stack(a,b)

19. A B C C A B Goal state Goal interaction • Simple planning algorithms assume that the goals to be achieved are independent • Each can be solved separately and then the solutions concatenated • This planning problem, called the “Sussman Anomaly,” is the classic example of the goal interaction problem: • Solving on(A,B) first (by doing unstack(C,A), stack(A,B) will be undone when solving the second goal on(B,C) (by doing unstack(A,B), stack(B,C)). • Classic STRIPS could not handle this, although minor modifications can get it to do simple cases Initial state

20. A B C C A B Sussman Anomaly Achieve on(a,b) via stack(a,b) with preconds: [holding(a),clear(b)] |Achieve holding(a) via pickup(a) with preconds: [ontable(a),clear(a),handempty] ||Achieve clear(a) via unstack(_1584,a) with preconds: [on(_1584,a),clear(_1584),handempty] ||Applying unstack(c,a) ||Achieve handempty via putdown(_2691) with preconds: [holding(_2691)] ||Applying putdown(c) |Applying pickup(a) Applying stack(a,b) Achieve on(b,c) via stack(b,c) with preconds: [holding(b),clear(c)] |Achieve holding(b) via pickup(b) with preconds: [ontable(b),clear(b),handempty] ||Achieve clear(b) via unstack(_5625,b) with preconds: [on(_5625,b),clear(_5625),handempty] ||Applying unstack(a,b) ||Achieve handempty via putdown(_6648) with preconds: [holding(_6648)] ||Applying putdown(a) |Applying pickup(b) Applying stack(b,c) Achieve on(a,b) via stack(a,b) with preconds: [holding(a),clear(b)] |Achieve holding(a) via pickup(a) with preconds: [ontable(a),clear(a),handempty] |Applying pickup(a) Applying stack(a,b) From [clear(b),clear(c),ontable(a),ontable(b),on(c,a),handempty] To [on(a,b),on(b,c),ontable(c)] Do: unstack(c,a) putdown(c) pickup(a) stack(a,b) unstack(a,b) putdown(a) pickup(b) stack(b,c) pickup(a) stack(a,b) Goal state Initial state

21. Algorithms for Finding Plans in State Space • STRIPS is intuitive, but has problems because it is ‘working as it thinks’ • Let’s think of another method… • Planning algorithms have a start-state • Actions, if applicable, will change the state • Goal: easily tested by analyzing the current state for the given conditions • Sounds like a general search problem! • Why we wasted the time with PDDL: better heuristics

22. Forward Search • Start with initial state, apply any operators for which the preconditions are satisfied • Repeat on frontier nodes until you reach the goal • Problem: Lots of wasted time exploring irrelevant actions • Problem: Schema may be small, but adding more fluents exponentially increases the size of the state space • Not all hope is lost: heuristics will be very helpful!

23. Backward (Regression) Relevant-States Search • Similar in philosophy to STRIPS • Start with goal, search backward to start state • ONLY explore actions that are relevant to satisfying the preconditions of the current state • Difference: search algorithms and heuristics can prevent ‘sussman’s anomaly’ • Backward search requires reversible actions, but PDDL is good for that • Also needs to allow state SETS (since the goal may have non-ground fluents) • Unification of variables allows backward search to dramatically reduce branching factor vs. forward search • Set-based heuristics are harder to design.

24. Planning Heuristics • Key advantage of PDDL and planning as its own topic: effective problem-independent heuristics • Insight: PDDL problem descriptions all have a similar structure, independent of problem area • Makes forward search for many planning problems feasible • Common choices • Ignore preconditions • Ignore delete lists • Subgoal independence • Pattern Databases

25. A new view on heuristics • We’ve already discussed viewing search in a state space as a graph search problem: • One way of thinking about designing a heuristic is to add edges to the state space graph • Shorter paths mean shorter (admissible) estimates for cost to the goal

26. Ignoring Preconditions • Apply actions regardless of whether the state has satisfied the preconditions • Value of heuristic: minimum number of actions such that the union of their effects-sets equals the set of conditions in the goal • Sound Familiar? • Approximations are effective, but lose A* admissibility requirement • May also ignore only selected preconditions • But this is a domain-dependent heuristic

27. Ignoring Delete Lists • Assume goals and preconditions only have positive literals (easy to augment any problem—how?) • What happens then? • Remove delete lists • Executing one action never ‘undoes’ another • Applying any action monotonically increases toward the goal • FF Planner (Hoffman 2005) uses this heuristic with a modified hill climbing search that keeps track of solution path

28. Subgoal Independence • Assume that solving each of K items in the goal is independent of the others • Run K planning searches, one for each goal then sum the cost of each sub-plan to estimate cost of total plan • Optimistic Heuristic: when negative interactions occur between subplans (one undoes the other) • Admissible • Pessimistic Heuristic: when subplans contain redundant actions • Not admissible

29. Pattern Databases • Frequently Called Plan Libraries • Many planners are used repeatedly using new start states • Action schemas make this possible • Record exact solutions to plans and sub-plans as you solve them • Research questions: • Detecting frequently recurring and useful sub-plans • Using plan libraries to solve problems in a different domain • based on qualitative graph structures (Bulka, 2009)

30. Planning Graphs • Planning Graph: secondary data structure generated from a planning problem • Analyzing this data structure leads to a very effective heuristic • Approximates a fully expanded search tree using polynomial space • Estimates the steps to reach the goal • Always correct if goal not reachable • Always underestimates (so what?)

31. Planning Graph Basics • A planning Graph is divided into levels • Two types of alternating levels: • State levels Si fluents that MIGHT be true at level i • Always underestimates the time at which it will actually be true • Action levels AiActions that MIGHT be applicable at step i • Takes into consideration some but not all negative interactions between actions • Negative interaction = Performing one action violates the preconditions of the other • “The level j at which a literal first appears is a good estimate of how difficult it is to achieve the literal from the initial state [in the final plan].”

32. Example Problem and Planning Graph Initial State: Have(Cake) Goal State: Have(Cake)  Eaten(Cake) Action(Bake(Cake) Pre: Have(Cake) Effect: Have(Cake)) Action(Eat(Cake) Pre: Have(Cake) Effect: Have(Cake)Eaten(Cake))

33. Constructing a Planning Graph • All action schemas must be propositionalized • Generate all possible grounded actions so no variables are left. • Start with S0 = all initially true fluents • Construct Ai = all actions whose preconditions are satisfied by Si • Construct Si = all fluents made true by the effects of the actions in Ai-1 • All levels in Aihave the NO-OPaction which passes all true fluents in Si to Si+1

34. Constructing a Planning Graph (Continued) • Add links: • Between levels • From fluents in Si to preconditions of actions in Ai • From effects of actions in Ai to fluents in Si+1 • Within Levels (mutual exclusion) • Action Mutex Link: Two actions compete for resources • State Mutex Link: Two fluents that cannot both be true at the same time

35. Rules for Constructing Mutex Links • Action Mutex: • Inconsistent Effects: One action negates the effect of the other (e.g., Eat(Cake) and Bake(Cake) ) • Interference: One of the effects of one action is the negation of a precondition of the other (e.g., Eat(Cake) and the persistence of Have(Cake)) • Competing Needs: One of the preconditions of one action is mutually exclusive with a precondition of the other • Depends on State Mutex links in previous level • State Mutex: • One is the negation of the other • All pairs of actions that could make both true have Acton mutex links between them

36. Example Problem and Planning Graph Initial State: Have(Cake) Goal State: Have(Cake)  Eaten(Cake) Action(Bake(Cake) Pre: Have(Cake) Effect: Have(Cake)) Action(Eat(Cake) Pre: Have(Cake) Effect: Have(Cake)Eaten(Cake))

37. Growth of the Planning Graph • Levels are added until two consecutive levels are identical (graph has leveled off) • How Big is this Graph? • For a (propositionalized) planning problem with L literals and a actions, • Each Si level: Max L nodes, L2mutex links • Each Ai level: Max a + L nodes, (a + L)2mutex links • Inter-level linkage: L(a + L) links • Graph with n levels  O(n(a+L)2)

38. Planning Graph Heuristics • If any goal literal fails appear, then the plan is not solvable • For a single goal: • Level at which it appears is a simple estimate for cost of achieving goal • Serial Planning Graphs insist that one action is performed at each level by adding mutex links between all non-NO-OP actions • Leads to more effective heuristic • For a conjunction of goals: next slide

39. Heuristics for Conjunction of Goals • Max-Level: Take the max of all goal literals • Admissible, but poorly conservative • Level-Sum: sum all goal costs together • not admissible but works OK much of the time • Set-Level: First level where all goal literals appear without Mutex links • Admissible, dominates Max-Level

40. Comments on Planning Graphs • Planning Graphs do have shortcomings: • Only guarantees failure in the obvious case • If goal does appear at some level, we can only say that there is ‘possibly’ a plan • In other words, there’s no obvious reason the plan should fail • Obvious reason = mutex relations • We could expand the process to consider higher-order mutexes • More accurate heuristics • Not usually worth the higher computational cost • Similar to arc-consistency in CSP’s

41. The Graph Plan Algorithm

42. Extracting a (Possible) Solution • Implement as Backward-search • States = A level of the graph, and a set of goals • Start at level Sn, add all main goal fluents • Select a set of non-mutex actions that achieve the goals • Also, their preconditions cannot be mutex • New state is Level Sn-1 with its goals being the preconditions of the chosen action set • If we reach S0 with all goals satisfied, Success! • Heuristic: Solve the highest-level fluents first, using the lowest cost action (based on preconditions)

43. About No-Goods • no-goods is a hash table: • Store (level,goals) pairs • If Extract-solution fails it will: • Record its current goals and level in no-goods • If Extract-solution reaches the exact same state later, we can stop early and return FAILURE (or, backtrack to a different part of search) • No-Goods is important for termination • Can’t just say ‘planning graph leveled off so we’re done’ • Planning graphs grow vertically faster than horizontally • Some plans need more horizontal space to be solved • Extract Solutions is thorough • If a plan exists, it will find it • If it can’t find a new reason to fail at level Si then adding a level won’t help

44. Other Approaches • Boolean satisfiability • Propositionalize actions, goal • Add successor-state axioms • Add pre-condition axioms • Use SATPLAN algorithm to find a solution • Seems complex, but is very fast in practice • Situation Calculus • Similar to satisfiability above • Incoporates all of FOL, more expressive than PDDL • Partial Order Planners (Up Next!)