Playing to Par: Reinforcement Learning for Provably Optimal Quadrilateral Block Decompositions
The agent is trained to hit the topology-imposed minimum for irregular vertices in all-quadrilateral meshes.
The paper defines that minimum as “par,” derived from the discrete Gauss-Bonnet identity. Its reinforcement-learning agent edits a mesh directly and uses mesh-connectivity convolutions so the policy can run on larger domains than it saw in training. On 96 held-out domains, it completed every mesh and reached provable optimality on 90. Gmsh, at the same element count, completed 51 and reached optimality on none. HF Daily Papers' note
The paper defines that minimum as “par,” derived from the discrete Gauss-Bonnet identity. Its reinforcement-learning agent edits a mesh directly and uses mesh-connectivity convolutions so the policy can run on larger domains than it saw in training. On 96 held-out domains, it completed every mesh and reached provable optimality on 90. Gmsh, at the same element count, completed 51 and reached optimality on none. HF Daily Papers' note
score 4