Russell and Norvig, Chapter 11: Planning

11.1 A Simple Planning Agent [Fig11.1, p338]
- compare to problem-solving agent [Fig3.1] and KB Agent [Fig6.1]
- we want to build Ideal-Planner

11.2 From Problem-Solving to Planning
- difficulties in problem-solving (i.e., search)
  - actions are black-box, state-next state procedures
    - do not describe what they can achieve
    - no subgoaling
  - states are complete
    - cannot function with partial information
  - plans considered only as unbroken sequences
    - no partial plans or subplans
    - most conjunctive goals are independent
  - no ability to work on intermediate goals directly
    - e.g., to get the gold, I should do this next

11.3 Planning in Situation Calculus
- initial state
  - at(agent,[1,1],s0) and not hasgold(agent,s0) and alive(agent,s0)
- goal state
  - exists(S) hasgold(agent,S) and at(agent,[1,1],s0) and alive(agent,S)
- operators
  - successor-state axioms
  - hasgold(agent,result(A,S)) <=>
      [(A = grab) and at(agent,[X,Y],S) and gold(X,Y)] V
      [(A != release) and hasgold(agent,S)]
- result'(L,S): state resulting from applying the sequence of actions in L
                starting at state S
- can use this to do planning, but not very practical (ie, theorem proving)
- using a planner
  - gives more guidance and is therefore more efficient
  - at the expense of a less-expressive representation
  - each representation requires a different planner

11.4 Basic Representations for Planning
- STRIPS (STanford Research Institute Problem Solver) language
- states and goals
  - states: conjunction of function-free ground literals
  - goals: conjunction of function-free literals
  - queries look for sequences of actions to obtain, not just truth
  - closed-world assumption is the default
- actions (operators)
  - name
  - precondition: conjunction of positive literals
  - effect: conjunction of literals
  - e.g., name: goforward0
          precondition: at(X,Y) and orient(0) [and not wall(X+1,Y)]
          effect: not at(X,Y) and at(X+1,Y)
- situation space vs. plan space
  - searching in situation space
    - progression: start from initial state (problem solving)
    - regression: start from goal state
  - searching in plan space
    - partial plan: start anywhere
      - refinement operators: add and/or order steps
    - least commitment strategy (variables and order)
  - usually fewer plans than situations
- plans: partial order or total order (--linearization-->)
  - set of plan steps
    - intially: { S1: start([],InitialState)
                  S2: finish(GoalState,[])
  - set of step ordering constraints, Si < Sj
    - initially: S1 < S2
  - set of variable binding constraints, v=x
    - initially: null
  - set of causal links (protection intervals)
    - causes(Si,C,Sj): Si achieves C for Sj
    - initially: null
  - e.g., [Fig11.5, p348]

11.5 A Partial-Order Planning Example [Fig11.6 - Fig11.12]
- protecting causal links by demotion or promotion [Fig11.10, p353]

11.6 A Partial-Order Planning Algorithm [Fig11.13]
- POP
- non-deterministic definition via 'choose' construct (Prolog)
- select_subgoal not a candidate for backtracking
  - order in which goals selected only affects efficiency

11.7 Planning with Partially-Instantiated Operators
- possible threats: not at(wumpus,X,Y) and at(wumpus,2,2)
  - (1) resolve now, binding X,Y to something other than 2,2
  - (2) resolve now, adding constraint that not(X=Y=2)
  - (3) resolve later, wait until X,Y bound by other constraints
    - implementation [Fig11.14, p358]
      - avoid nested loops in resolve_threats by storing dependencies

11.8 Knowledge Engineering for Planning
- blocks world
- Shakey's world
- wumpus
  - operators correspond to actions (e.g., goforward, turnleft, ...)

11.9 Summary
- closed-world assumption and the wumpus world (incomplete information)?
- conditional effects (e.g., shoot)?
- the above concerns answered in Chapter 12