Workshop on Planning and Inference

News

• The program includes Keynote talks, invited presentations and contributed papers .
• Complete schedule posted below.
• Overview slides posted.

Overview

The workshop is focused on the problems of Stochastic Planning and Probabilistic Inference and the intimate connections between them. Both Planning and inference are core tasks in AI and the connections between them have been long recognized. However, much of the work in these subareas is disjoint. The last decade has seen many exciting developments with explicit constructions and reductions between planning and inference that aim for efficient algorithms for large scale problems and applications. The work in this area is is distributed across many conferences, sub-communities, and sub-topics and varies from discrete to continuous problems, single vs. multi-agent problems, general vs. spatial problems, propositional vs. relational problems, model based planning vs. reinforcement learning, and exact/optimal vs. approximate vs. heuristic solutions. Applications similarly vary for example from scheduling to sustainability and to robot control.

The goal of this workshop is to bring together researchers from all these areas and facilitate synergy and exchange of ideas: to discuss core ideas, techniques and algorithms that take advantage of the connection between planning and inference, identify opportunities and challenges for future work, and explore applications and how they can inform the development of such work.

The workshop will include invited talks by experts on planning and inference, contributed talks and a poster session, leaving room for discussion and interaction among participants.

The workshop topic is broad and the intention of this first workshop is to enable interaction among the various sub-areas while keeping the focus on the interaction between planning and inference. Some basic questions:

• What are effective reductions from planning to inference?
• What are effective inference algorithms for such problems?
• What are the challenges in planning applications, and how does their structure help or interfere with the application of planning as inference?
• Can generic inference algorithms be used directly for planning? or are we better off tailoring algorithms directly to the planning problem?
• Can planning algorithms or ideas developed for them be used for general inference?
• How do structured solutions, e.g., lifted inference, lifted planning, spatial MDPs, cooperative multi-agent systems, and approximations in continuous problems, translate across the planning/inference spectrum, and help improve scalability.
• Success stories and challenges in using planning for inference or inference for planning.
• These questions cut across theoretical foundations and practical applications.

Paper Submission

We invite 4 types of submissions (typeset in the AAAI style):

• Papers describing current unpublished work (up to 8 pages including references).
• Review of mature work (from multiple papers) by the authors (up to 8 pages including references).
• Papers recently published at other venues (1 page abstract with a link to the full paper).
• Position papers (2 pages including references).

Submissions of papers being reviewed for AAAI 2018, or at other venues are welcome since this is a non archival venue (and if published they can be replaced with a 1 page abstract). If such papers are currently under blind review, please anonymize the submission.

Submission procedure: Please submit your paper through https://easychair.org/conferences/?conf=planinf2018

Important Dates

• Friday, October 13, 2017: Paper submission deadline.
  Electronic papers due by 11:59 PM UTC-10 (midnight Hawaii)
• Thursday, November 9, 2017: Notifications Sent to Authors
• Tuesday, November 21, 2017: Final Workshop Papers Due at AAAI
• Workshop date: Friday, February 2, 2018

Conference Committee

Organizers

• Roni Khardon, Tufts University, USA
• Akshat Kumar, Singapore Management University, Singapore
• Alex Ihler, UC Irvine, USA

Program Committee

• Christopher Amato
• David Barber
• Vaishak Belle
• Alan Fern
• Kristian Kersting
• Qiang Liu
• Radu Marinescu
• Denis Maua
• Sriraam Natarajan
• Gerhard Neumann
• Pascal Poupart
• David Poole
• Regis Sabbadin
• Scott Sanner
• Prasad Tadepalli
• Jan-Willem van de Meent

Workshop Program

8:30-10:30 - Session 1

8:30- 8:45 Introductory notes and Workshop overview

8:45- 9:30 - Keynote Talk: Rina Dechter: On Search solvers for Marginal Map and their applicability to Probabilistic Planning.

9:30- 9:50 - Generalized Dual Decomposition for Bounding Maximum Expected Utility of Influence Diagrams with Perfect Recall. Junkyu Lee, Alexander Ihler and Rina Dechter.

9:50-10:10 - Lifted Stochastic Planning, Belief Propagation and Marginal MAP. Hao Cui and Roni Khardon.

10:10-10:30 - R2PG: Risk-Sensitive and Reliable Policy Gradient. Bo Liu, Ji Liu and Kenan Xiao.

10:30-11:00 - Coffee Break

11:00-12:30 - Session 2

11:00-11:30 - Invited Presentation: David Wingate Probabilistic Programming for Theory of Mind for Autonomous Agents.

11:30-12:00 - Invited Presentation: Jan-Willem van de Meent: Probabilistic programming for planning as inference problems.

12:00-12:30 - Invited Presentation: Pascal Poupart: Planning as Marginal MAP and Stochastic SAT. [abstract]

Abstract: In this talk, I will explain how various decision theoretic planning algorithms can be reduced to stochastic satisfiability (SAT) and marginal maximum a posteriori (MAP) inference problems. I will also describe stochastic SAT algorithms and hybrid message passing techniques that can be used for planning in stochastic and partially observable environments.

12:30-2:00 - Lunch Break (lunch on your own)

2:00-3:30 - Session 3

2:00- 2:45 - Keynote Talk: Pascal Van Hentenryck: Planning for Energy and Transportation Systems. (cancelled due to travel delay)

2:45- 3:05 - Learning Others' Intentional Models in Multi-Agent Settings Using Interactive POMDPs. Yanlin Han and Piotr Gmytrasiewicz.

3:05- 3:25 - Planning and Learning For Decentralized MDPs With Event Driven Rewards. Tarun Gupta, Akshat Kumar and Praveen Paruchuri.

3:30-4:00 - Coffee Break

4:00-5:30 - Session 4

4:00- 4:45 - Keynote Talk: Marc Toussaint: Physical Manipulation Planning and Sufficient Symbols.

4:45- 5:15 - Invited Presentation: Qiang Liu: Reasoning and Decisions in High Dimensions -- A Unified Approach. [abstract]

Abstract: Probabilistic graphical models such as Markov random fields, Bayesian networks and decision networks (a.k.a. influence diagrams) provide one of the most powerful frameworks for representing and exploiting dependence structures of high dimensional variables. The last two decades have witnessed significant improvements in basic inference tasks on graphical models, including combinatorial optimization and marginalization (e.g., computing data likelihoods or probabilities of evidence), particularly based on variational inference algorithms. However, modern machine learning applications increasingly require more challenging inference tasks, such as robust combinatorial optimization with missing information or latent variables, stochastic programming and decision making in single- or multi-agent systems. These problems are notoriously challenging: they require both optimizing over large numbers of decision variables and averaging over random variables, in either simultaneous or sequential environments.
In this talk, I present a unified variational representation for all these problems, providing a broad and powerful framework for deriving new classes of efficient exact or approximate algorithms. In particular, I describe a class of "message-passing" style algorithms that are simple, fast and nicely amenable to parallel or distributed computation. We show that our algorithms significantly outperform earlier algorithms both in terms of empirical results and theoretical guarantees.

5:15- 5:30 - Conclusion / Discussion.

Contact Information