Profile cover photo
Profile photo
Christian Muise
441 followers -
Research Fellow at the University of Melbourne
Research Fellow at the University of Melbourne

441 followers
About
Christian's interests
View all
Christian's posts

Post has attachment
This was the view from my dinner table tonight. #hlf15
Photo

Post has attachment
Some pics from the trip to Beijing for IJCAI 2013.
PhotoPhotoPhotoPhotoPhoto
IJCAI 2013
55 Photos - View album

Hey all,
  It looks as though I'll be putting in a final push to finish a first draft of my thesis over the next few weeks, and so won't have time to run another MM until after ICAPS. Hopefully most of us will be able to meet up there, and we'll have lots to discuss afterwards I'm sure.

  Cheers,
   Christian

Google just blew my mind. These searches all returned exactly what you might hope from my own personal photos:
- "my photos of kristie" (my wife +Kristie Taylor-Muise)
- "my photos of sunsets" (only a few had actual captions)
- "my photos of music" (less accurate, but again it nailed some photos that had no caption -- just a group of musicians playing live on stage)
- "my photos of my cat" (I don't have photos of other cats, but I doubt Google is /that/ smart...right?)


Problem Generation as Planning

For posterity sake, I just wanted to recap our previous discussion briefly. The modelling issue considered was how to generate planning problems or domains through the use of automated planning technologies. As far as I can recall, we considered two options.

1. Anti STRIPS
In this approach, the domain and initial state is assumed to be given, and we're looking for a "challenging goal" for a planner to achieve. The name comes from the STRIPS heuristic, which is just a count on how many fluents from the goal do not hold in the current state.

The actions are stripped of their cost, and we set the goal to be the all-False state (assuming we're in a STRIPS domain as well). A special phase-shift action is introduced that allows the planner to stop executing regular actions, and execute newly introduced actions that simply delete a single fluent. Any plan will be some sequence of actions from the original problem, followed by the "phase-shift" action, and then finally the fluent deleting actions. A cost-optimal plan will minimize the cost of fluent deleting actions.

The final piece of the puzzle is to set the cost of the fluent deleting actions. In the first iteration, every action that deletes a fluent from the initial state has a cost of 1. After the first plan is found, the sequence of regular actions brings us to a new state. For every fluent in that new state, we increase the corresponding fluent deleting action cost. This would (in theory) cause diverse goals to be generated after every iteration. Pick a cutoff, and take the reached state from the final iteration.

The criticism of this approach is that just as the STRIPS heuristic is a horrible estimate of the true plan cost, the actual difficulty of a planning problem is likely not modelled directly by this diversification approach.


2. Heuristic Dive
The second idea we considered was tailoring a planning solver to randomly roll out to a "deep" state (high distance from the initial state). The general idea is to use enforced hill climbing, but instead of evaluating a heuristic for state s as h(s,G), we evaluate it as h(s,I) and take the maximum (not minimum). Intuition here is that the planner would be directed as far from the initial state as possible.

We also tossed around the idea of biasing this approach to take paths that have bad heuristic guesses (i.e., drastically underestimate the true distance to goal). This could be done by computing the next step based on the discrepancy between an expensive / accurate heuristic compared to a cheap / rough one.


3. Other?
If I've missed anything, let me know in the comments.

Any suggestions on how to repair a collection in Mendeley given a proper .bib file? The bib file has page numbers for all of the articles (70 or so), but the numbers have been accidentally removed from the Mendeley version. I'd rather not just re-import the correct bib file because it would lose all of the pdf attachments (I presume).

  Any options?

Hey everyone. Just wanted to remind you all about the MM session tomorrow -- it's been delayed a few times so you may not have received a notification.

+Amanda Coles / +Nir Lipovetzky: Welcome to the group -- feel free to join us (2pm UTC) if you're around. We cap the discussion to an hour, so it doesn't take out too much of the day.

Post has attachment
Yipes...

Keeping with the bi-weekly schedule set up (carefully designed to avoid key dates in the future ;)), the next MM will be a week from tomorrow (March 4th). We'll be checking out how some planners cope with previously discussed encodings. See you then.

Hey all. Tomorrow is a holiday here in Toronto, and unfortunately I won't be around for MM. Hope to fire things up again next week. Cheers
Wait while more posts are being loaded