Probabilistic Planning vs. Replanning. Little, Thiebaux. ICAPS Workshop on International Planning Competition: Past, Present and Future.

[Reading this because of a comment on the HOLOP paper]

  1. Overall, a bit confusingly written, and makes some points I’m not so confident about
  2. Say that planning in stochastic domains is good when there is noise and the problem isn’t terribly hard.  Deterministic planning (with replanning on failure) is better for cases when the noise is minimal, or the domains are too large to reason about probabilistically
  3. I don’t understand why, but say probabilistic planning is best in closed-loop cases when behavior must occur quickly (I suppose because the contingency planning already happened, so you dont have to stop and sit there while replanning happens, as deterministic planners do?)
  4. Replanning is done online, while probabilistic planning is usually done offline (I think for the same basic reason I gave above)
  5. Success of FF-Replan in planning competition helped put this conventional wisdom on its head
  6. Contribution of the paper is to give a test for when planning is accomplished most efficiently by replanning, so by extension when to do probabilistic planning and when to do determistic (re)planning
  7. Gist of the paper seems to be that probabilistic planning has fallen out of favor, so the axe-to-frind is throwing some water on the deterministic replanning approach
  8. They say that in the planning competition, the domains are not “interesting” (which they want to be able to measure), because the smaller domains are optimally or near-optimally solvable by replanning, and the large domains are too big for probabilistic planning
  9. Results here are only relevant to planning problems where a goal state is sought.
  10. Their notation is silly.
  11. Discuss a couple of methods of “compiling” a stochastic domain to a deterministic one
  12. Define a few structural qualities that are important from a planning perspective:
    1. Dead end states – from which the goal is unreachable
    2. How avoidable dead end states are based on action selection
    3. Number of distinct trajectories possible from start to goal
    4. Mutual exclusion; choices that exclude other useful choices later on.
  13. Talk about avoidable and unavoidable dead ends.  An unavoidable dead end is one where there is a nonzero probability of reaching the dead end and, trying to work around it will reduce the probability of reaching the goal
    1. They say unavoidable dead ends aren’t interesting, because they say it is impossible to improve on the solution one finds with replanning (I don’t think this is true).
  14. In general, there are too many points that are made and not backed up so I’m leaving it alone here.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: