The 8th Workshop in the MSDM series
May 7, 2013
St. Paul, Minnesota, USA

http://gaips.inesc-id.pt/~switwicki/msdm2013/

Overview

In sequential decision making, an agent's objective is to choose actions, based on its observations of the world, in such a way that it expects to optimize its performance measure over the course of a series of such decisions. In environments where action consequences are non-deterministic or observations incomplete, Markov decision processes (MDPs) and partially observable MDPs (POMDPs) serve as the basis for principled approaches to single-agent sequential decision making. Extending these models to systems of multiple agents has become the subject of an increasingly active area of research over the past decade and a variety of models have emerged (e.g., MMDP, Dec-POMDP, MTDP, I-POMDP, and POSG). The high computational complexity of these models has driven researchers to develop multiagent planning and learning methods that exploit the structure present in agents' interactions, methods that provide efficient approximate solutions, and methods that distribute computation among the agents.

The MSDM workshop serves several purposes. The primary purpose is to bring together researchers in the field of MSDM to present and discuss new work and preliminary ideas. Moreover, we aim to identify recent trends, to establish important directions for future research, and to discuss some of the topics mentioned below such as challenging application areas (e.g., cooperative robotics, distributed sensor and/or communication networks, decision support systems) and suitable evaluation methodologies. Lastly, MSDM seeks to bring researchers from other AAMAS communities together in order to facilitate consensus among different models and methods, thus making the field more accessible to new researchers and practitioners.

Topics

Multiagent sequential decision making comprises (1) problem representation, (2) planning, (3) coordination, and (4) learning. The MSDM workshop addresses this full range of aspects. Topics of particular interest include:

• Challenging conventional assumptions, e.g,
  • model specification: where do the models come from?
  • how should we decide an appropriate level of abstraction for our model descriptions?
• Novel representations, algorithms and complexity results
• Comparisons of algorithms
• Relationships between models and their assumptions
• Decentralized vs. centralized planning approaches
• Online vs. offline planning
• Communication and coordination during execution
• Computational issues involving...
  • large numbers of agents
  • large numbers of / continuous states, observations and actions
  • long decision horizons
• (Reinforcement) learning in partially observable multiagent systems
• Cooperative, competitive, and self-interested agents
• Application domains
• Benchmarks and evaluation methodologies
• Standardization of software
• High-level principles in MSDM: past trends and future directions

Invited Speaker

Dr. Kobi Gal (Ben-Gurion University)

Panel

This year, we are excited to be hosting a joint panel of the MSDM and ALA worshops. We have confirmed the participation of Shlomo Zilberstein, Peter Stone, Milind Tambe, and Sandip Sen.

Accepted Papers

Planning under Uncertainty for Coordinating Infrastructural Maintenance
Joris Scharpff, Matthijs T.J. Spaan, Leentje Volker and Mathijs de Weerdt

Asynchronous Execution in Multiagent POMDPs: Reasoning over Partially-Observable Events
João V. Messias, Matthijs T.J. Spaan and Pedro U. Lima

Organizational Design Principles and Techniques for Decision-Theoretic Agents
Jason Sleight and Edmund H. Durfee

Rehearsal Based Multi-agent Reinforcement Learning of Decentralized Plans
Landon Kraemer and Bikramjit Banerjee

Counterfactual Regret Minimization for Decentralized Planning
Bikramjit Banerjee and Landon Kraemer

Automated Generation of Interaction Graphs for Value-Factored Decentralized POMDPs
William Yeoh, Akshat Kumar and Shlomo Zilberstein

Opponent modeling and planning against non-stationary strategies
Pablo Hernandez-Leal, Enrique Muñoz de Cote and L. Enrique Sucar

Qualitative Planning under Partial Observability in Multi-Agent Domains
Ronen I. Brafman, Guy Shani and Shlomo Zilberstein

Solving Dec-POMDPs by Genetic Algorithms: Robot Soccer Case Study
Okan Aşik and H. Levent Akin

A Coordinated MDP Approach to Multi-Agent Planning for Resource Allocation, with Applications to Healthcare
Hadi Hosseini, Jesse Hoey and Robin Cohen

Bayesian Reinforcement Learning for Multiagent Systems with State Uncertainty
Christopher Amato and Frans A. Oliehoek

Organizing Committee

Program Committee

Sponsors

This workshop is supported, in part, by the Portuguese Fundação para a Ciéncia e Tecnologia (FCT) and by Carnegie Mellon Portugal Program and its Information and Communications Technologies Institute, under project CMU-PT/SIA/0023/2009 ("MAIS-S"), and by the FCT (INESC-ID multiannual funding) under project PEst-OE/EEI/LA0021/2011.