To assess the effect of native support for ADL constructs on the performance of Marvin, we ran a series of tests comparing the planner's performance when given both the STRIPS and ADL domain encodings. Macro-actions were disabled in both cases to isolate the effect the encoding itself was having on performance. In IPC 4, ADL was used to encode four of the domains: Airport, Philosophers, Optical Telegraph and PSR. STRIPS compilations were made available for each of these domains, in which each ground action that could arise when using the original ADL domain was made into a fixed-parameter STRIPS action. In the Philosophers, Optical Telegraph and PSR domains, the domain formulations making use of Derived Predicates were used.
In the Airport, Optical Telegraph and PSR domains, the performance of Marvin (with macro-actions disabled) was unaffected by the use of either the ADL or STRIPS domain encoding. The ADL domain encodings did not give rise to inefficient compiled STRIPS encodings.
In the Philosophers domain, the use of the ADL domain encoding resulted in a reduction in planning time when compared to the use of the compiled STRIPS encoding. As can be seen in Figure 17, more problems can be solved within the 30 minute time-limit if the ADL encoding rather than the STRIPS encoding is used, even disregarding the improvements in performance provided by the use of macro-actions.
Andrew Coles and Amanda Smith 2007-01-09