From: Tal Shaked (tshaked_at_u.washington.edu)
Date: Fri May 16 2003 - 00:07:00 PDT
Leap Before You Look: Information Gathering in the PUCCINI planner – Golden
This paper discusses how to soundly reason about actions that have
conditional effects (which may or may not be known to be satisfied at
plan/execution time) in the framework of a partial order planner by
introducing the idea of verification links.
Actions have preconditions that must be satisfied for an action to be
applicable at a certain state, as well as conditional effects which have
causes if certain conditions (secondary conditions) are satisfied. There
are cases when an action is applied and it is unknown if the secondary
conditions are met. Later in the plan these secondary conditions may become
known and therefore the effects (which at the time of execution were
unknown) may be inferred and used as conditions for goals or other actions.
Verification links represent these dependencies and protect the conditions.
Specifically an action that produces a secondary condition can occur before
the action that uses it (in which case the agent immediately can reason
about the conditional effect), be the action itself, or occur after the
action that uses it which is the more interesting (and complicated) case.
Although this adds extra complexity to the search space since the planner
has more ordering options, it provides more possibilities and therefore can
achieve a larger set goals (and conceivably more quickly with the extra
flexibility). Furthermore, empirical results seem to indicate that the
extra complexity does not make any significant difference in computation
time.
One problem with just about every paper that talks about Puccini is that it
is unclear how the ideas fit in a planner that interleaves planning with
execution. Verification links can connect with sensing actions so even
though a plan can be sound after being fully executed, there are
contingencies along the way that complicate the matter, and more discussion
of how this mixes with execution might help.
Related to the above, I believe there is room to improve how an agent
decides when to execute a partial (or complete branch of a) contingent plan
(perhaps up to a verification link), and when to continue planning, looking
for a ‘better’ solution. No heuristics were discussed in how to deal with
verification links such as where to place the actions (perhaps trying to
keep them close to the related action), or when to choose them from the set
of flaws, but these may be quite useful.
This archive was generated by hypermail 2.1.6 : Fri May 16 2003 - 00:07:17 PDT