Learning Heuristics for A* Search (demo paper)
A learned heuristic guides A* to the goal under a tight expansion budget.
Tier 1 · Understand
The paper replaces hand-tuned search heuristics with a learned estimate of cost-to-go, letting A* expand far fewer nodes while still finding near-optimal paths. Naive uninformed search (BFS) exhausts its budget on large maps; the learned heuristic focuses expansion toward the goal.
- 1Searching a maze blindly wastes moves exploring dead ends.
- 2A heuristic is a smart guess of how far the goal still is.
- 3The paper learns that guess from data instead of hand-coding it.
- 4Result: the same A* algorithm reaches the goal with a fraction of the steps.
Concept map
Ask the paper
💬 Ask the paper — answers are grounded in its text, with sections cited
Checking…
Tier 2 & 3 · Play / Prove
template: pathfinding · conf 92%🎯 Goal — Reach the goal tile before the step budget runs out.
Budget exhausted after 60 expansions — no path found.
The maze represents a planning graph where each cell is a state in the search space. Blind BFS wastes its budget exploring dead-end states.
Toggle to the paper's heuristic to see it focus the search.
Paper claim: learned heuristic reaches the goal using ~5x fewer expansions than BFS.