Advancements in POMDP Solvers
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.
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.
The workshop will begin with
a set of tutorials, followed by paper presentations. Paper presentations will
be 20 minutes long, including questions.
8:30 Workshop starts - presenting the agenda
8:50 – 10:30 Tutorials
8:50 - 9:10 POMDP basics
9:10 – 9:50 Point-based algorithms
9:50 - 10:30 Online POMDP Solvers
10:30 - 10:45 break
10:45 – 12:00 Tutorials
10:45 - 11:20 Policy Iteration and Finite State
Controllers
11:20 - 12:00 Factored POMDPs and
Structured Representations
12:00 - 12:20 Applications
Hand washing – Pascal Poupart
Assisting
elderly people – Joelle Pineau
Spoken
dialog – Jason Williams
13:50 - 14:30 Technical session I:
Hierarchical POMDP Controller
Optimization by Likelihood Maximization – Marc Toussaint, Laurent Charlin, Pascal Poupart
Optimizing Fixed-Size
Stochastic Controllers for POMDPs - Christopher Amato, Daniel S. Bernstein,
Shlomo Zilberstein
14:30 - 15:30 Technical session II:
POMDP and DEC-POMDP
Point-Based Observation Aggregation - Alan Carlin, Shlomo Zilberstein
Point-Based Value Iteration
Algorithms for Multi-Core Machines – Guy Shani
15:30 - 15:45 break
15:45 - 16:45 Technical session III:
Solving POMDPs from Both
Sides: Growing Dual Parsimonious Bounds - Nicholas Armstrong-Crews, Geoffrey
Gordon, Manuela Veloso
Continuous-State POMDPs with
Hybrid Dynamics - Emma Brunskill, Leslie Kaelbling,
Tomas Lozano-Perez, Nicholas Roy
16:45 - 17:30 Technical session IV:
Cooperative active perception
using POMDPs – Matthijs Spaan
Integrating expert knowledge
into POMDP optimization for spoken dialog systems – Jason Williams
17:30-17:45 - closing remarks
·
Joelle Pineau, McGill University, jpineau@cs.mcgill.ca
·
Pascal Poupart, University of
Waterloo, ppoupart@cs.uwaterloo.ca
·
Guy Shani, Microsoft Research, guyshani@microsoft.com, 425-7051561
·
Trey Smith, Carnegie Mellon
University West / NASA Ames Research Center, trey.smith@west.cmu.edu
·
Call for participation - doc, pdf.
·
AAAI 2008 webpage - http://www.aaai.org/Conferences/AAAI/aaai08.php.
·
AAAI paper format guidelines - http://www.aaai.org/Publications/Author/author.php.
·
ICAPS 2008 Uncertainty Planning
Track - http://ippc-2008.loria.fr/wiki/index.php/Main_Page.