Chapter 7: Gathering Insights

Abstract

We have seen in Part II, specifically in Chapter 3, that most of the planning benchmarks can very efficiently be solved by a rather simple algorithm: the FF base architecture, using an h ⁺ approximation in a greedy local search strategy, dramatically outperforms all previous approaches across almost all of our domains, c.f. Section 3.7. This opens up the question, why does such a simple approach work so well in so many domains? What are the underlying common patterns of structure to these domains? Can we find a characterization of the domains where the approach works well?