Artificial Intelligence - Planning
What is a blocks world in the context of planning?
It is a micro-world consisting of a robot hand, table, and set of blocks using logic notation.
What is the Sussman Anomaly?
It is the behavior of simple planning algorithms to undo sub-goals if they conflict with each other. Classic Strips had this with the assumption that once a goal had been satisfied, it would stay satisfied.
How does hierarchical task network planning differ from the other planning algorithms?
It starts from a high-level abstraction of a goal and decomposes it into lower-level abstractions. It continues this until it gets to primitive operators.
Is situational calculus efficient?
No! It is exponential in worst case! It also doesn't guarantee a good plan. Use Planning!
Situation calculus planning
Representing planning problem using first-order logic. Lets us reason about changes in the world.
What planning approach does Shakey the Robot use?
STRIPS! (Stanford Research Institute Problem Solver)
What does STRIPS seek?
Seeks a sequence of actions to attain goal, not seeking if goal is true.
What are the two additional data structures data structures maintained by STRIPS?
State List and Goal Stack
What are the delete-effects of an operator?
The literals that are negative after the operators.
What are add-effects of an operator?
The literals that are positive after the operators.
What do refinement operators do?
They add constraints to the partial plan and modification operators for other changes.
How are states represented in situational calculus?
They are represented by logical sentences about the state. Example: At(Home, S0 ) Ù ¬Have(Milk, S0 ) Ù ¬ Have(Bananas, S0 ) Ù ¬Have(Drill, S0 )
How do non-linear planners embody the principle of least commitment?
They only add actions, orderings, and variable bindings to the plan that are absolutely necessary to reach the goal state.
What are regression planners?
They search backward from initial state to goal state. Needs operators to have information for both ways and has difficulty handling conjunctions of goals.
What are Progression planners?
They search forward from initial state to goal state. Usually result in a high branching factor.
How do simple planning algorithms use the assumption of independent sub-goals?
They solve each sub-goal separately and concatenate the solution.
How can a partially ordered plan be linearized?
Through a topological sort.
How can we do planning as search?
We generate successor states and our goal is represented as a goal state using a heuristic function. We find a sequence of states that gets us to the goal state either forward or backward. Not very efficient.
How do we fix the Sussman Anomaly?
We need the notion of protecting a sub-goal so it is not undone by later steps.
How do planning and problem solving compare?
While they solve similar problems, planning is more powerful and efficient. Planning often goes through the plan space rather than state space.
What do you do if the current goal is not satisfied by the present state?
You find the operator that adds it (makes it true) and push that operator then its preconditions onto the stack.
What is the planning problem?
Finding the sequence of actions to achieve a goal state when executed from initial state given the set of possible primitive actions, start state description, and goal state description.
What do non-linear planners do?
Build plans as set of steps with temporal constraints.
How are goals represented in STRIPS?
A conjunction of ground literals, but may have variables assumed to be existentially quantified.
How are states represented in STRIPS?
A conjunction of ground literals.
What is a precondition in an action operator?
A conjunction of positive literals that must be true to execute the operator.
What is the effect in an action operator?
A conjunction of positive or negative literals that describe the situation after the operator is applied
What is the State List?
A list of all currently true predicates.
What is the goal stack?
A push down stack of goals that need to be solved with the current goal on top.
General Problem Solver
A system that was an early planner. Introduced general way to use domain knowledge to select most promising action to take next.
What are the components of an action operators?
Action description, preconditions, and effect.
How do you refine a new partially ordered plan?
Adding a new plan step or adding constraints to the steps already in the plan.
