Connectionists: Call for papers: AAAI 2008 Workshop on Advancements in POMDP Solvers

[Joelle Pineau]

AAAI 2008 Workshop on Advancements in POMDP Solvers
http://research.microsoft.com/~guyshani/AAAIWorkshop/Main.html

Description:
Over the past decade, much advancement was achieved in the field of POMDP
solvers.
The size of POMDPs that solvers can handle has increased by orders of
magnitude.
Solvers developed ten years ago were hardly able to handle more than 10
states,
while modern solvers scale up to models with millions of states. New
techniques
compute approximate policies of manageable complexity, thus allowing us to
handle
larger and more complicated POMDPs. This advancement was achieved by a few
orthogonal approaches - the use of point-based techniques, finite-state
controllers,
efficient model representations, model compression techniques, hierarchical
decompositions, inference-based techniques, and improved algorithms for
online
search.

Topics:
This workshop is designed to bring together researchers currently working
on POMDP
solvers, as well as those with complimentary interests. A main objective
is to
identify remaining barriers preventing full-scale POMDP deployment, as
well as key
scientific opportunities. The workshop is also intended to introduce new
people to
the current advancements in POMDP solvers.

Format:
The workshop will begin with a set of short tutorials over various
advancements in
POMDP solvers:
                1.Point based algorithms - insights into the point-based
approach.
Overview of the basic operations and the various approaches.
Introduction to relevant algorithms: PBVI, Perseus, HSVI, FSVI.
                2.Compression and efficient representation - an overview
of various
techniques for compressing POMDPs, such as belief state compression
and linear compression. A description of efficient descriptions for
factored POMDPs such as Algebraic Decision Diagrams (ADDs).
                3.Online belief space search - overview of various
algorithms and
new advancements in online belief space search.
                4.Policy iteration algorithms - using techniques such as
finite
state controllers to efficiently learn policies directly, without
computing a value function.
The workshop will continue by a set of short presentations of work in the
field of
POMDP solvers. Accepted papers will have 15 minutes for presentation of
the main
contributions. We will then follow up by a poster presentation of all
accepted
papers that will allow for an extended discussion with the authors of the
papers.

Submissions:
We invite researchers in the field to submit papers discussing:
                *New methods for solving POMDPs.
                *Improvements of existing solvers.
                *Empirical studies that illuminate the difficulties and
opportunities in POMDP solvers.
                *Complimentary methods for optimal control under partial
observability.
                *Studies of implementations of POMDP solvers to real life
problems.
Submissions should be 6 pages long and follow the standard AAAI guidelines
(http://www.aaai.org/Publications/Author/author.php).
We also welcome submissions of papers that were previously published in
conferences
and journals.
Authors of accepted papers must sign the distribution permission form
(http://www.aaai.org/Publications/Author/distribute-permission.pdf).
A PDF or PS version of the paper should be submitted by email to
guyshani at microsoft.com.

Schedule:
                *Submission deadline: April 7, 2008
                *Notification date: April 21, 2008
                *Accepted paper submission deadline: May 5, 2008

Organizing Committee:
                *Joelle Pineau, McGill University, jpineau at cs.mcgill.ca
                *Pascal Poupart, University of Waterloo,
ppoupart at cs.uwaterloo.ca
                *Guy Shani,  Microsoft Research, guyshani at microsoft.com,
425-7051561
                *Trey Smith, Carnegie Mellon University West / NASA Ames
Research
Center, trey.smith at west.cmu.edu