Workshop on Planning and Learning

Workshop Program

9:00-10:30	Invited Talk: Abstraction and Learning for Probabilistic Planning Leslie Pack Kaelbling Learning to Combine Admissible Heuristics Under Bounded Time Carmel Domshlak, Erez Karpas and Shaul Markovitch
10:30-11:00	Coffee Break
11:00-12:15	Learning Instance-Specific Macros Maher Alhossaini and J. Christopher Beck Learning action effects in partially observable domains Kira Mourão, Ronald P. A. Petrick and Mark Steedman Planning with the help of Statistical Relational Learning Ingo Thon, Bernd Gutmann, Martijn van Otterlo, Niels Landwehr and Luc De Raedt
12:15-14:00	Lunch
14:00-15:30	Exploiting N-gram Analysis to Predict Operator Sequences Christian Muise, Sheila McIlraith, Jorge A. Baier and Michael Reimer Learning Divide-and-Evolve Parameter Configurations with Racing Jacques Bibaï, Pierre Savéant, Marc Schoenauer and Vincent Vidal Learning Weighted Rule Sets for Forward Search Planning Yuehua Xu, Alan Fern and Sungwook Yoon Three Relational Learning Approaches for Lookahead Heuristic Planning Tomás de la Rosa, Sergio Jiménez, Rocío García-Durán, Fernando Fernández, Angel García-Olaya,Daniel Borrajo
15:30-16:00	Coffee Break
16:00-17:30	Learning and Exploiting Configuration Knowledge for a Portfolio-based planner Alfonso E. Gerevini, Alessandro Saetti and Mauro Vallati Discussion Panel

Workshop Description

Progress in planning research faces the perennial battle of how to provide effective search guidance, marred further by the aim to produce domain-independent systems with heuristics and algorithms that are effective in many classes of problems. Great progress has been made in these areas, and several successful approaches have emerged - useful domain-independent heuristics, landmarks analysis, novel propagators for use in a planning-as-CSP framework, to name but a few. Nonetheless, despite their elegance, they are invariably out-performed by systems able to exploit hand-crafted domain-specific search guidance. The drawback of such guidance, however, is the amount of human time needed to produce suitable guidance for each domain - the key motivation behind domain-independent approaches.

Between domain-independent search and hand-crafted domains-specific guidance lies the pragmatic stance of applying learning techniques to planning. The aim is to eliminate the human bottleneck by automating the process of determining domains-specific guidance. In doing so, the system as a whole becomes domain independent once again - the learning system can be used on each domain for which search guidance is desired. Assuming the learning time is reasonable for the task in hand, the system can then be used to find solutions exploiting this guidance without the user ever having to consider how it might be obtained.

In the past few years, learning in planning has re-emerged as a vibrant research area. Following the 2007 ICAPS workshop on Planning and Learning, a 'Learning Track' was organised for the subsequent 2008 International Planning Competition (IPC). The IPC tested the efficacy of the participating learning systems, producing an interesting collection of results.

This workshop provides a forum for discussing issues surrounding the use of learning techniques in planning, continuing the lineage of the events of ICAPS 2007 and 2008. The topics that will be covered include, but are not limited to:

• Approaches to learning search guidance
• Representation of learnt knowledge - control rules, heuristics, macro-actions...
• Applying learning to portfolio-based planners
• Hybrid learnt-guidance--generic-heuristic search
• Learning for optimal planning
• Applications of planning and learning
• Lessons learnt from the 2008 IPC
• Future Challenges for the IPC Learning Track

We invite contributions from researchers who have considered the application of learning to planning. Successful approaches are welcome, as well as negative results from which the community may learn. Further, details of many of the systems participating in the 2008 competition have yet to be published, and are of great community interest, so we welcome details of these systems. We also welcome theoretical contributions considering the expressive power and/or limitations of various forms of learnt knowledge representation. Additionally, we also welcome ideas for solving planning problems based around presently hand-crafted rules that may be amenable to learning approaches.

Important Dates

Submission Deadline:	June 23rd. June 28th, 2009 (Included)
Notifications and Technical Program:	July 22nd, 2009
Camera-Ready Paper Submissions:	August 3rd, 2009 (Included)
Edited Working Notes:	August 9th, 2009
Workshop Date:	September 20th, 2009

Submission Procedure

Submissions may be full papers (8 pages) or short papers (2 pages). Full papers should report work in progress or completed work. Authors of full papers that are accepted by the Programme Committee will be invited to give a talk on the paper. Short papers papers should report views or ambitions, describe problems or applications. Authors of short papers will be invited to give a short talk on their work and will be able to discuss the paper informally with others at the workshop.

Please note that all submitted papers will be carefully peer-reviewed by multiple reviewers, and that low-quality or off-topic papers will not be accepted. Also note that all workshop participants must register for the main ICAPS-09 conference

We ask authors to submit papers in PDF format. Papers should be formatted in accordance with the AAAI style template and may be at most 8 pages long, including figures and bibliography. Visit the url http://www.aaai.org/Publications/Author/author.php for formatting instructions.

Papers can be submitted via e-mail or made available on the Web. In either case, documents should be in gzipped PDF format and be named "author.pdf.gz", using the name of the first author. An e-mail message containing either the file or its URL (e.g. http://..../author.pdf.gz) should reach Andrew Coles () by June 23rd 2009.

Workshop Participation

As well as dissemination of research through presentations on accepted papers, we aim to provide a forum for the discussion and debate of current issues surrounding learning and planning. To this end, the workshop will consist of discussion sessions where audience members are welcome (indeed, encouraged) to participate. As a further stimulus to this, the organisers will provide a commentary following each session of talks.

Program Committee

• Daniel Borrajo, Universidad Carlos III de Madrid, Spain
• Adi Botea, NICTA/ANU, Australia
• Alan Fern, Oregon State University, USA
• Alfonso Gerevini, Universitá di Brescia, Italy
• John Levine, University of Strathclyde, UK
• N.S. Narayanaswamy, Indian Institute of Technology Madras, India
• Sungwook Yoon, Arizona State University, USA

Organizing Commitee

• Amanda Coles. University of Strathclyde, Glasgow, UK
• Andrew Coles. University of Strathclyde, Glasgow, UK
• Susana Fernández. Universidad Carlos III de Madrid, SPAIN
• Sergio Jiménez. Universidad Carlos III de Madrid, SPAIN
• Tomás de la Rosa. Universidad Carlos III de Madrid, SPAIN

University of StrathClyde	Universidad Carlos III de Madrid