Heuristic search is a standard technique for planning, but it can be memory-intensive. The paper notes that even a nearly perfect heuristic does not prevent the search from storing exponentially many states.
To address this, the authors propose learning search control rather than relying solely on heuristic evaluation. They learn one specification per domain, expressed as an indexical policy—a generalized policy that guides the search.
This shifts the focus from improving heuristics to acquiring control knowledge that reduces the space needed during planning. The paper presents this as a way to make search dramatically more space-efficient.