The macro-action strategy adopted by Marvin in IPC 4 was to generate its macro-actions on a per-problem basis. It is possible, however, to build libraries of macro-actions on a per-domain basis; this approach was taken by Macro-FF [Botea et al. 2005]. Marvin's macro-actions could also be cached for use when solving all the problem instances in a given domain. If this were done, then the knowledge encapsulated in the plateau-escaping macro-actions that allows heuristic imperfections in the search landscape to be bypassed could be made available across all the problems in a given domain without needing exhaustive search to re-discover this knowledge on each problem instance. In contrast to existing systems that use off-line learning to generate and test macro-actions, caching Marvin's plateau-escaping macro-actions across solving a problem suite in this manner would allow for online learning to take place. Further work is being undertaken in this area, to investigate effective caching strategies and to manage the large number of macro-actions found.
The idea of using plateau-escaping macro-actions is not restricted to search under the relaxed planning graph heuristic. Currently, the effect of using the macro-actions in search under other heuristics is being investigated, including the causal-graph heuristic [Helmert 2004] used by Fast-Downward .
At present, the macro-actions used in Marvin are restricted to those used to escape plateaux. Work is currently in progress exploring ways of extending the macro-learning capabilities of Marvin to include more general macro-action structures of the kind being explored by Botea and Schaeffer [Botea et al. 2005].
Andrew Coles and Amanda Smith 2007-01-09