Wizard's Home
Good News! Wizard clearly worn in 2 out of 6 domains in efficiency and obtained 2nd position in learning for improvement but 3rd and 4th position in overall performance among 13 participant systems in the 6th International Planning Competition 2008 Learning Track. Wizard is a macro-learning method and its co-participant planners were Metric-FF and SGPlan. Note, in the competition, Wizard learnt macros that help obtain faster planning (i.e. efficiency), no matter how they affect plan-length (i.e. quality). Please click here to get the poster on the results.
- Wizard is a learning method that acquires macro-actions, or macros, with a view to improving automated planning by knowledge assistance. Given a planner, a domain, and a few example problems, Wizard suggests macros that help the planner solve future problems in the domain faster. However, it has been extended to cope with improving a linear combination of plan-time, plan-length, and plan-metric.
- Wizard's effort is focused and confined only within macro acquisition; for representation of macros and their exploitation during planning, Wizard relies on the support available. Currently planners do not reason about macros. Wizard therefore compiles macros into actions and adds them into the domain. The additional actions in effect produce a transformed search space, which might appear more accessible to the planner.
- Unlike existing work, Wizard does not explicitly discover or exploit any specific planner or domain properties. It therefore works with arbitrary planners and domains, where arbitrariness is in the characteristics borne or exhibited. Wizard is suitable when improving performance is the objective but any specific characteristics are not known or are of no concern.
- Wizard learns individual macros by exploring the entire macro space. Thus, it virtually learns any types of macros. This means Wizard can learn macros that are not observable from the given examples, that are not learnt by any existing methods, and that have action orderings not explored by the planner. Wizard also explores the macro-set space to learn collections of macros that maximise the performance by interacting among themselves.
- Wizard adopts an evolutionary method to generate macros using actions lifted from generalised plans of small example problems. It further employs a sophisticated procedure to evaluate them by solving other large example problems. The combination of unified search space formulations, a guided exploration approach, a carefully designed evaluation strategy, and their overall trial-and-error nature has made Wizard an effective and generalised macro acquisition method.
- Is Wizard a planner or a learner? Wizard is a learner. It is not a planner at all. Wizard learns macro-actions for given domain-planner pairs. It does not focus on any particular planner or domain characteristics.
- What PDDL subsets does Wizard support? Wizard currently works with propositional (STRIPS) and numerical (NUMERICAL-FLUENTS) domains only; this is because neither PDDL nor the planners support macros. We have a plan to extend Wizard to cope with functional domains (OBJECT-FLUENTS).
- Does Wizard consider plan-length or plan-metric? Although initially Wizard was designed to improve plan-time, its current distribution does consider a linear combination of plan-time, plan-length, and plan-metric.
- Does Wizard learn macros that improve a number of planners on a domain? Currently Wizard learns macros that are specific to a given planner on a given domain. However, we do have a plan to extend it in this dimension.
- Wizard: Learning Macro-Actions Comprehensively for Planning; M. A. H. Newton; supervised by J. Levine, and M. Fox; Doctor of Philosophy Thesis, Department of Computer and Information Science, University of Strathclyde, United Kingdom; November 2008. [thesis (onesided)] [thesis (twosided)] [presentation] [highlight]
- Learning Macros That Are Not Captured by Given Example Plans; M. A. H. Newton, J. Levine, M. Fox, and D. Long; Supplementary Proceedings for Poster Papers at the 18th International Conference on Automated Planning and Scheduling (ICAPS); September 2008. [paper] [poster]
- Wizard: Compiled Macro-Actions for Planner-Domain Pairs; M. A. H. Newton, J. Levine, M. Fox, and D. Long; Booklet for the 6th International Planning Competition Learning Track; September 2008. [paper]
- Learning Macro-Actions for Arbitrary Planners and Domains; M. A. H. Newton, J. Levine, M. Fox, and D. Long; Proceedings of the 17th International Conference on Automated Planning and Scheduling (ICAPS); September 2007; nominated for the best paper award. [paper] [presentation]
- Wizard: Suggesting Macro-Actions Comprehensively; M. A. H. Newton and J. Levine; Proceedings of the Doctoral Consortium held at the 17th International Conference on Automated Planning and Scheduling (ICAPS); September 2007. [paper] [presentation] [poster]
- Evolving Macro-Actions for Planning; M. A. H. Newton and J. Levine; Proceedings of the Workshop on AI Planning and Learning held at the 17th International Conference on Automated Planning and Scheduling (ICAPS); September 2007. [paper] [presentation] [poster]
- Combinations of Domain Enhancing Macro-Actions in Planning; M. A. H. Newton and J. Levine; Proceedings of the Workshop on General Artificial Intelligence held at the 13th Portuguese Conference on Artificial Intelligence (EPIA); December 2007. [paper] [presentation]
- Learning Macro-Actions Genetically from Plans; M. A. H. Newton, J. Levine, M. Fox, and D. Long; Proceedings of the 25th Workshop of the UK Planning and Scheduling Special Interest Group (PlanSIG); December 2006. [paper]
- Genetically Evolved Macro-Actions in AI Planning Problems; M. A. H. Newton, J. Levine, and M. Fox; Proceedings of the 24th Workshop of the UK Planning and Scheduling Special Interest Group (PlanSIG); December 2005. [paper]
Please email M.A.H. Newton for any query or bug reporting.
| Release Date |
Version Name |
Description |
03.09.2008 |
wizard.pn |
This is the first complete release version that works with propositional and numeric constructs and learns macros to improve plan-time, plan-length, plan-metric and/or their linear combinations. It has a user manual (as part of the executable) and sample files. |
| 03.09.2008 |
wizard.ipc |
This version was submitted to the 6th International Planning Competition, 2008, Learning Track. It does not come with sufficient user instructions and is therefore not recommended for use.
|
| 03.09.2008 |
wizard.phd |
This version was used in the experiments of my PhD thesis. It does not come with sufficient user instructions and is therefore not recommended for use.
|
Supported by the Commonwealth Scholarship Commission in the United Kingdom under the Commonwealth Scholarship & Fellowship Plan (CSFP).
M.A.H.Newton 13.01.2009