Maps & Abstractability

Experiments use a set of hand‑specified gridworlds at sizes 3×3, 5×5, and 9×9, defined in config.yml under the game key. The top‑left tile is the start, the bottom‑right tile is the goal, and X marks obstacles that block movement. The Rust core treats these maps deterministically and exposes state counts and enumerations to Python.

We characterize how “abstractable” a map is with the reduction factor R = (n − k) / n, where n is the number of reachable ground states and k is the number of abstract states in the ideal abstraction derived by the Rust core. Following the thresholds in llm_abstraction/utils/classify.py, we label maps as “perfect abstraction” when R ≥ 0.25, “partial abstraction” when 0 < R < 0.25, and “no abstraction” when R = 0. These labels provide a quick sense of difficulty before running planners (Thesis).

Symmetry matters: layouts with symmetric obstacle patterns tend to admit higher‑quality abstractions and yield better planning gains for all agents. The preview-maps command writes thumbnails to outputs/maps/<map_hash>/, which helps spot symmetries and anticipate results.