Planning with Temporally Extended Preferences and Goals

**UNDER CONSTRUCTION**

**Until this page is updated, feel free to email me for details regarding software and/or papers. See also my publications page.

Researchers

Sheila McIlraith

Research Description

Research in automated planning has historically focused on classical planning -- generating a sequence of actions to achieve a user-defined (final state) goal, given a specification of a domain and an initial state. Nevertheless, in many real-world settings, plans are plentiful, and it is the generation of high-quality plans meeting users' preferences and constraints that presents the biggest challenge. Our group has been doing a lot of work on planning with preferences.

Qualitative temporal preference language and optimal planner, PPLAN: We have developed a language for describing rich first-order qualitative temporally-extended preferences. We have also developed a k-optimal planner, called PPLAN, that plans in this language using an interesting optimistic-pessimistic best-first search heuristic. The best paper to read on this subject is the KR06 paper by Bienvenu, Fritz and McIlraith. Earlier work appeared in 2005 symposium and workshop proceedings. See the PPLAN web page for details, including papers, code and a nice test domain.
Agent programming + qualitative and quantitative preferences: We have developed a means of integrating qualitative and quantitative preferences into agent programming. In particular, we devised a way of integrating qualitative preferences into DT-Golog. The best reference for this work is the KR06 paper by Fritz and McIlraith. Earlier work appeared in NRAC2005.
A domain-independent compilation algorithm and a heuristic search planner: Informed by the work on PPLAN, we were motivated to develop a preference planner that more actively worked towards achievements of user preferences. To that end, we first worked on planning with temporally extended goals (TEGs) using heuristic search. The best reference to this work is a AAAI06 paper by Baier and McIlraith . Most interesting in this work is a compiler that compiles a planning problem with TEGs into a classical planning problem. The compiler is available online and can be used with any classical planner. Experimental results often showed orders of magnitude speedup over existing techniques.
We then decided to exploit some of these ideas to develop a heuristic search planner for temporally extended preferences (TEPs). In particular, we decided to integrate our ideas for planning with preferences into Bacchus and Kabanza's TLPlan and to enter it into the International Planning Competition (IPC-5). We modified our compiler to accept PDDL3.1 TEPs and TEGs, developed representations for TLPlan, develped a set of domain-independent heuristics for planning with TEPs that we evaluated against IPC-5 benchmarks. Our planner showed very good performance, ranking 2nd in the qualitative preferences track behind SGPlan5, whose planning efficieincy (without preferences) dominated the competition. We have short workshop paper by Baier, Hussell, Bacchus and McIlraith describing this work and a draft paper under review.
There is other work in the pipeline, not yet published..... There is also interesting related work on planning with temporally extended goals.