Bratko, Chapter 16: Planning

16.1 Representing Actions
- preconditions: can(Action,Cond)
- add-list: add(Action,AddRels)
- delete-list: del(Action,DelRels)
- e.g., move(Block,From,To) [Fig16.2, p402]
        camera world [Fig16.3, p403]

16.2 Deriving Plans by Means-Ends Analysis
- plan(State,Goals,Plan,FinalState) [Fig16.5, p406]
- assumes all actions and goals fully instantiated

16.3 Protecting Goals
- e.g. state: on(b,3); goals: clear(2), clear(3); move(b,3,2) ...
- plan(State,Goals,ProtectedGoals,Plan,FinalState) [Fig16.6, p411]

16.4 Procedural Aspects and Breadth-First Regime
- PrePlan is determined using breadth-first because of conc
  - conc(PrePlan,_,_) generates PrePlan = [], [_], [_,_], ...
- rest of plan generate using depth-first
- can move to breadth-first by adding conc(Plan,_,_) to plan/4
- means-ends planning is incomplete (assumes goal independence)
- Exercise 16.4 relates to HW3

16.5 Goal Regression [Fig16.8, p417]
- given Action and Goals (none of which are deleted by Action, and
  at least one of which is added by Action)
- find Goals0 = Goals - addlist(Action) + preconds(Action)
- allows for goal interaction

16.6 Combining Means-Ends Planning with Best-First Heuristic
- can add domain knowledge
  - select(State,Goals,Goal)
  - achieves(Action,Goal)
  - which alternative regressed goal sets to attempt next
- best-first search based on number of unachieved goals [Fig19.9, p420]
  - states are goal sets and an action
  - initial state is desired goals
  - goal state is any goal set satisfied by the initial facts
  - successor implemented using regression (all costs = 1)
  - bestfirst(Goals -> stop, Plan)

16.7 Further Refinements: Uninstantiated Actions and Non-Linear Planning
- uninstantiated goals and actions
  - would like to have variables in goals
  - problems maintaining constraints when variables are not instantiated
    - e.g., enforcing Object != To in move(Object,From,To)
  - check for constraints during goal testing
- non-linear planning
  - produce partial-order plans