Russell and Norvig, Chapter 12: Practical Planning

12.1 Practical Planners
- OPTIMUM-AIV: spacecraft assembly, integration and verification
  - replanning after execution fails to meet expectations
  - extensions to Strips operator representation
    - hierarchical plans
    - complex conditions
      - universal quantification
      - conditional effects
    - time: projects have deadlines, actions have duration
    - resources: budget and manpower
  - uses Open PLANning (O-PLAN) architecture
    - similar to POP
- job shop scheduling
  - various products, machines and processes
- scheduling for space missions
  - e.g., processing hubble requests
    - SPIKE: long-term scheduling by week
    - SPSS: short-term planning
  - similar to the job-shop scheduling problem
- buildings, aircraft carriers and beer factories
  - System for Interactive Planning and Execution monitoring (SIPE)
    - deductive rules over states
    - object/constraint inheritance hierarchy

12.2 Hierarchical Decomposition
- consider action shoot vs the low-level operations for a robot to do so
- decompose shoot as grab arrow, grab bow, position arrow, draw bow, aim,
  release ; each of which can be further decomposed
- extending Strips language
  - operators are primitive or non-primitive
  - non-primitive operators contain decomposition plans
  - plan must be consistent
    - no ordering or variable inconsistencies
  - every effect of operator must be asserted by at least one step of plan
  - every precondition of the steps in plan must be asserted by a step in
    the plan or be a precondition of the operator
- hierarchical plans allow a way to store previously computed plans
- hierarchical decomposition planner (HD-POP) [Fig12.2, p374]
  - loops until solution, each time selecting a subgoal and a decomposition
  - for decomposition, modify bindings, orderings and links

12.3 Analysis of Hierarchical Decomposition
- hierarchical planning can be better than non-hierarchical planning if
  the following properties hold
  - downward solution property: if p is an abstract solution, then there is
    a primitive solution of which p is an abstract
    - can prune other abstract solutions
  - upward solution property: if an abstract plan is inconsistent, then
    there is no primitive solution of which it is an abstraction
    - can prune all descendants of any inconsistent abstract plan
- otherwise, hier. planning has same worst-case performance as nonhier.
- decomposition and sharing
  - two decompositions of two steps of an abstract plan may need to share
    steps
  - when adding operators, consider sharing as an alternative choice point
  - or, go ahead and merge operators, but let a critic detect the need to
    share
- decomposition vs approximation
  - an operator can be represented by an approximation hierarchy in which
    preconditions are assigned a criticality level
  - planning can occur at a high criticality level (low value) first, then
    at less critical levels
  - motivation: solve most highly constrained preconditions first

12.4 More Expressive Operator Descriptions
- conditional effects
  - e.g., did shooting the arrow kill the wumpus?
  - effects now look like "e [when c]"
  - consider c as a subgoal only when e is protected by a causal link
  - given a causal link Si --c--> Sj and an effect ((not c') when p),
    ensure p does not hold when c unifies with c'
    - this is called confrontation [Fig12.7, p382]
- negated and disjunctive goals
  - prove (not p) using an operator with (not p) effect, or p not true in
    initial state
  - backtracking handles which of (p or q) is used in a precondition
    - could split operator into two, put forces committment
  - disjunctive effects harder to handle, see Chapter 13
    - environment becomes non-deterministic
- universal quantification
  - operator: CarryBag
    - preconditions: clear(b) becomes forall(X) block(X) -> not on(X,b)
    - effects: forall(I) item(I) -> (at(I,Y) and not at(I,X)) when in(I,bag)
  - this is really enumeration, i.e., static universe
    - forall goals can be expanded to long conjunction
    - forall effects can be used as needed (for threats or links)
- Partial-Order Planning with Disjunction, Universal Quantification,
  Negation and Conditional effects (POP-DUNC) [Fig12.8, p385]

12.5 Resource Constraints
- use measures: gallons(5), $(1.50), etc
  - declared up front with constraints (e.g., legal ranges)
- conditions and effects allow inequality, arithmetic and assignment
- time can be treated as a resource