Transcript Planning in AI

PLANNING IN AI
PLANNING IN AI
• Determine the set of steps that are necessary to
achieve a goal
• Some steps might be conditional, i.e., they are only
taken when a set of conditions is present during
plan execution.
TYPICAL PLANNING PROBLEMS *
• Representing and reasoning about time, causality,
and intentions
• Physical and other kinds of constraints on
acceptable solutions
• Uncertainty in the execution of plans
• How the “real world” is sensed and perceived
• Multiple agents who may cooperate or interfere
* Lucci, Stephen; Kopec, Danny: Artificial Intelligence in the 21st
Century
TERMINOLOGY
• Operator Schema
• The sequence of steps followed to solve a planning
problem
• Characterize actions/events
STRIPS
(STANFORD RESEARCH INSTITUE PROBLEM SOLVER)
• Operator Schema Consist of:
•
•
•
•
Precondition
Delete List
Add List
Variables in the Precondition, Delete List, and Add List that
are bound at run-time
• Example:
•
•
•
•
Pickup(X)
Precondition: OnTable(X) ^ HandEmpty ^ Clear(X)
Delete List: OnTable(X) HandEmpty Clear(X)
Add List: Holding(X)
SOME TYPES OF PLANNING
• Hierarchical Planning
• Goals are not equally valued
• Some might be necessary
• Others might be desirable, but not necessary
• Numerical values provide hierarchical value
• Opportunistic Planning
• Exploit the conditions in a plan state to more easily achieve
a goal
• Conditional Planning
• Planning based on things that “might happen.”
• Partial Order Planning
• Not all steps in the plan must be made in order
PLANNING STRATEGIES
• Means-Ends Analysis
• Reduce the distance between the current state and the
goal state (GPS – Newell and Simon)
• Least Commitment Planner
• Only commit to a plan when forced by constraints
• Can maintain multiple/flexible “parallel” plans until forced
to select a concrete alternative
• Ordering of steps is deferred as long as possible
• Depth-First Backtracking (“lifting”)
• Generates combinations of steps until one that works is
found
• Impractical for large plans (exponential time)
PLANNING STRATEGIES
(CONTINUED)
• Beam Search
• Expand the top few (beam width) nodes at each level of a
breadth-first search
• One-The-Best Backtracking
• Rely on local information to determine a best-guess path,
then backtrack – looking for good alternative candidates
along the path
• Dependency-Directed Search
• Store dependencies between decisions, assumptions, and
alternatives for each
• On failure, keep parts of the solution that are not
dependent of the cause of failure
PLANNING STRATEGIES
(CONTINUED)
• Opportunistic Search
• Favor the most constrained operations
• Meta-level Plans
• Plans about plans
• Techniques of planning can be selected dynamically based on
the type of problem at hand.
• Distributed Planning
• Different parts (subplans) of the plan are developed by
different “expert” components of the system
GRAPHPLAN
• Graph builds a graph starting from the initial state
• Uses “layered” or “parallel” plans, where each layer provides a
set of actions that can be performed simultaneously.
• Mutually exclusive actions are identified at the level where
they first occur.
• Builds a graph where each node represents the state after all
possible transitions from the previous state.
• Solutions are found by working backward from the final node
• GraphPlan can also be used to generate heuristics for A*
search
• http://web.engr.oregonstate.edu/~afern/classes/cs533/
notes/graphplan.pdf