Second, macros can improve the heuristic evaluation of states (the evaluation effect). As shown before, FF computes this heuristic by solving a relaxed planning problem (i.e., the delete effects of actions are ignored) in a graphplan framework. To illustrate the benefits of macros in relaxed graphplan, consider the example in Figures 6 and 7. Operator UNLOAD has one add effect (LIFTING) and one delete effect (AVAILABLE) that update the status of a hoist from available (free) to lifting (busy). Similarly, operator DROP updates the hoist status with two such effects. However, when macro UNLOAD--DROP is used, the status of the hoist does not change: it was available (free) before, it will be available after. No effects are necessary to express changes in the hoist status. Hence two delete effects (one for each operator) are safely eliminated from the real problem before relaxation is performed. The relaxed problem is more similar to the real problem and the information loss is less drastic. See Section 5.4 for an empirical evaluation of how macros added to a domain affect the heuristic state evaluation with relaxed graphplan.
When macros can be added to the original domain formulation, both the evaluation effect and the embedding effect are present, with no need to extend the original planning engine. The disadvantages of this alternative include the limitation to STRIPS domains and, often, a significant increase of the preprocessing costs, memory requirements, or cost per node in the search (as shown in Section 5). When SOL-EP macros are used, each of the two effects needs a special extension of the planning engine. The current implementation of the enhanced planner handles the embedding effects but does not affect the computation of the heuristic state evaluation. Improving the heuristic state evaluation with macros is an important topic for future work.