MEvo is able to dynamically define a suitable sets of macros given a pool of macros.

“

We see several perspectives for future work. We are interested in performing an analysis of smarter ways for generating the Random set of MEvo, and the use of online macro-generation approaches (e.g., (Armano et al., 2004; Coles et al., 2007)) for increasing the pool of considered macros over time. We plan to investigate techniques for estimating the appropriate size of the different macro sets, instead of having a predefined maximum value of n = 3. We also envisage to explore more aggressive techniques for allowing MEvo to incorporate macros into domain models, for instance by removing original operators that macros incorporate: this can improve the performance by reducing the branching factor of problems, at the cost of the potential unsolvability of problems. We are also interested in the investigation of features-based approaches (see e.g., (Cenamor et al., 2013; Fawcett et al., 2014)) in order to extract more data for updating macro scores. Furthermore, we plan to investigate how the MEvo approach can be combined with approaches that aim at configuring the order in which macros and original operators are listed in the domain model (Vallati et al., 2015b). Finally, we plan to combine MEvo with approaches, such as ASAP or PbP (Gerevini et al., 2014; Vallati et al., 2014), that are also able to combine planners into a domain-specific portfolio. That would allow to select suitable planners, and evolve the domain model accordingly.

”