Submission: Kevin Joseph

[nadarenator]

Submission

• Leaderboard Position: 1
• Name: Kevin Joseph
• Overlap: 0.0000
• Wirelength: 0.2480
• Runtime: 435.52s

Solution: https://github.com/nadarenator/intern_challenge

Approach

3-phase gradient descent optimizer:

Phase 1 — Minimum wirelength initialization: Optimizes purely for wirelength, pulling each cell toward the centroid of its connections. This gives a strong starting layout biased toward short wire lengths, which carries through to the final result.

Phase 2 — Overlap elimination with exponential λ-annealing: Starts with a strong wirelength component (λ1=0.99) that preserves the phase 1 layout while overlap is eliminated. The overlap weight grows from 0.01 → 1 via an exponential schedule (k=19), ensuring large designs still converge.

Phase 3 — Wirelength fine-tuning with overlap guard: Pulls cells together to further reduce wirelength. The instant any overlap reappears, the last zero-overlap snapshot is restored and training stops. Empirically found to be redundant.