Russell and Norvig, Chapter 13: Planning and Acting

- planning with incomplete and/or incorrect information
  - new cave (wumpus? pits?) or moving wumpus
- conditional planning (contingency planning)
  - including contingencies in operator definitions
- execution monitoring
  - replanning when expectations are not met

13.1 Conditional Planning
- plan steps are now conditional on the current state of the world
  - if wumpus_ahead then goaround else goforward
- continue decomposing plan steps until non-conditional operator reached
  - conditional planning agent [Fig13.1, p395]
  - to ensure conditional plan is executable, agent must insert actions
    causing relevant conditions to be known
    - sensing actions whose effects are like: knows_whether(wumpus_ahead)
- generating conditional plans
  - the context of a plan step is the facts that must be true for that step
    to execute
    - e.g., wumpus_ahead or not wumpus_ahead
  - when a precondition cannot be met, and there is a sensing action for
    the precondition
    - let the sensing action support the precondition
    - form a new finish step whose context is the negation of the
      precondition (or in general the negation of the disjunction of
      previous contexts) 
  - Conditional Partial-Order Planning (CPOP) [Fig13.8, p399]
- parameterized plans
  - sensing actions may return several possible values
    - e.g., knows_what(color(X,C))
  - maintenance goals
    - e.g., maintain(color(X,C))
- similar to automatic programming
  - add a while construct
  - e.g., while((safe_ahead and not glitter), [goforward])

13.2 A Simple Replanning Agent
- methods for monitoring plans in unpredictable worlds
  - to each plan step, attach all preconditions of rest of plan
  - action monitoring, only check preconditions of next action
  - require good percepts
- inaccessible worlds, where relevant conditions are not perceivable, are
  beyond the scope of this chapter (uncertainty)
- causes of plan failure
  - bounded indeterminacy: possible failures can be enumerated
  - unbounded indeterminacy
- simple replanning agent [Fig13.9, p402]
  - if the preconditions of a plan segment p (of the entire plan q) are not
    met, choose the easiest (in terms of completing q) point p' in q and
    restart from there
- if all else fails, the agent will have to learn (see Chapter 21)

13.3 Fully Integrated Planning and Execution
- situated planning agent [Fig13.17, p408]
  - action, perception and deliberation are intertwined and concurrent
  - if an action's only purpose is the achieve a precondition, and that
    precondition becomes true, then the action can be removed
    - the removed action is called redundant
    - this process is called extending a causal link
  - actions are removed from the plan once executed
  - when no more actions in plan
    - start state directly supports the finish conditions, OR
    - more planning/acting is necessary to achieve failed expectations

13.4 Discussions and Extensions
- extensions
  - goal-oriented planning agent
  - partial-order planning
  - conditional effects
  - responds to changes in environment expectations
- conditional planning vs replanning
  - condition planning requires too many contingencies
  - replanning may not be prepared for contingencies
  - a tradeoff of the two is better
    - likely and/or costly contingencies are conditionally planned for
    - rare, less costly contingencies can be replanned
    - agent could learn conditional plan steps with experience
- abstraction
  - plan only at the level needed
  - leave the details for later