Learning Macro Actions for Artificial Intelligence Planning
J. Christopher Beck
The acquisition and use of macro actions has been shown to be effective in
improving the speed of AI planners. Current macro acquisition works
focus on finding one set of macro operators that, when added to the planning domain,
can improve the average performance of the planner on that domain.
In this project, we build a planner-independent, instance-specific domain remodeling
approach that uses macro operators and depends on machine learning.
We choose the best remodeling of the domain by building a predictor that
is based on previously solved problem instance using macros. Training the predictor
is achieved off-line, based on the observed correlation between the problem
instance features and the planner performance on macro-augmented domains. The
prediction of the best choice for remodeling the domain is achieved on-line
based on the problem instance features.
June 1, 2008
This work was partly supported by a Scholarship from King Saud University, Riyadh, Saudi Arabia.
It was also supported by a Fellowship from the University of Toronto.
- M. Alhossaini and J.C. Beck. Learning Instance-Specific Macros. In ICAPS 2009 Workshop on Planning and Learning, 2009.
- M. Alhossaini. Learning to Choose Instance-Specific Macro Operators. In ICAPS 2010 Doctoral Consortium, 2010.