class notes (powerpoint)
Download
Report
Transcript class notes (powerpoint)
Complexity of domain-independent planning
Jim Blythe
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
1
Reviewing decidability
A decision problem is a set of related problem instances,
e.g. “is 1 prime?”, “is 2 prime?” etc.
A decision problem is decidable if there is a program
that takes any instance and correctly halts with the
answer “yes” or “no”.
A problem is semi-decidable if a program guarantees to
halt and give the correct answer in one of the cases
[either “yes” or “no”] but not in the other case.
Otherwise, any program will sometimes run forever in
either case - this problem is undecidable.
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
2
Decidability results from Erol et al.
Consider plan existence, given operators, objects,
predicates an initial state and goal description.
Based on transformations to logic programming.
Can transform a planning problem without delete lists
or negative preconditions to a logic program (and
vice versa) in polynomial time.
function symbols => undecidable unless have
acyclicity and boundedness conditions.
No function symbols, and finite initial state =>
decidable.
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
3
Acyclicity
A “level mapping” maps propositions to natural
numbers. A “predicate level mapping” works on
predicate symbols, e.g. on, at-truck,..
If there is a level mapping such that for every
operator, l(add) > l(pre) for every literal in the add list
or preconditions, the P is atomically acyclic
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
4
Worst-case complexity of problems.
If a problem is decidable, we might ask how many
resources a program requires to compute the answer
(in the worst case).
We measure the resources a program takes in terms
of time or memory (space), as a function of the size
of the input.
If a problem is known to be in some complexity class,
then we know there is a program that solves it using
resources bound by that class.
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
5
Complexity classes
If a problem is in P, there exists a program to decide
it taking polynomial time in the size of the input.
PSPACE - polynomial space, EXPTIME and
EXPSPACE - exponential.
If a problem is in NP, there exists a nondeterministic
program that can solve it in polynomial time. You can
think of this program as one that guesses the correct
answer and then must verify it is correct.
Also have NEXPTIME, etc.
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
6
Hierarchy of complexity
We don’t know if there are any problems in NP but
not in P. But clearly NP contains P.
EXPSPACE
NEXPTIME
EXPTIME
PSPACE
NP
P
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
7
States, operators, plans.
How many, how big?
If there are n objects, m predicates with arity r and o
operators (with s variables each):
There are A = m * n^r possible atoms. It takes this much
space to describe a state.
So there are 2^A possible states (double exponential).
There are o * n^s possible ground operators.
In general plans will be bounded by the number of states.
(Why?)
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
8
Bounds on complexity
Now we’ll assume no function symbols and only
finitely many constant symbols, so the plan existence
problem is decidable.
With no restrictions, the problem is in EXPSPACE.
(What’s an algorithm that shows this?)
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
9
Some special cases - no delete lists.
If there are no delete lists, then operators only need
to appear once.
So it’s in NEXPTIME. (What’s an algorithm that
shows this?)
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
10
No negative preconditions
and no delete lists
Now plans for different subgoals won’t negatively
interfere with each other.
Now it’s in EXPTIME.
If each operator has only one precondition, the
number of subgoals does not increase in a
backward-chaining search - now it’s in PSPACE.
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
11
Propositional case
If the predicates have no arguments, so nor do operators,
the number of possible operators and atoms is polynomial.
This reduces the complexity of the algorithms we have
considered until now:
General case -> PSPACE
No delete lists -> NP
No negative preconds -> P
Not more than one precondition -> NLOGSPACE
If you know the operators in advance, this in effect bounds
the arity of predicates and operators, with the same result.
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
12
What does all this mean?
Domain-independent planning in general is very
hard.
The search for special cases seems fruitful.
e.g. Backstrom and Klein: each op changes at most
one proposition, no more than one op to make a
proposition true (or false). -> P.
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
13
Other approaches
Try to reuse past experience, with machine learning
or case-based reasoning. Even if a general problem
domain is hard, perhaps the subset we typically
encounter can be solved more easily and we can
cache problem parts that are reusable.
There are many possible approaches to this:
Learning search control knowledge
Learning macro-operators
Saving and reusing cases
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
14
We still need to think about each domain
These results paint a general landscape.
Even if a problem is in P, it can still be prohibitively
expensive to solve.
Many special cases exist that aren’t covered in this
framework.
What is the complexity of plan existence for the
blocksworld? PSPACE?
USC INFORMATION SCIENCES INSTITUTE
CS 541: Planning complexiity
15