What it does
Inspiration
Spatial scheduling of blocks is a real, costly bottleneck in shipyards. We wanted an approach that learns good placement patterns instead of relying only on fixed rules, so it adapts to the structure of each instance rather than treating every puzzle the same way.
What it does It solves the Grand Shipyard Puzzle (packing + scheduling) by proposing a strong initial layout with a learned model and then refining it under the time limit to maximize efficiency, always returning a valid solution.
How we built it A machine-learning model, trained on generated practice instances, guides the order and placement of blocks to produce a fast initial solution, which is then polished by local search (large-neighborhood search / simulated annealing) within the time budget. The model acts as a smart warm-start and classical optimization does the fine-tuning.
Challenges we ran into Training a model that generalizes across instances rather than overfitting to a few, and combining it with local search so the solution keeps improving without timing out ("Beat the Clock").
Accomplishments that we're proud of Designing a hybrid that pairs learning with classical optimization — using the model where intuition helps and exact-style search where precision matters — and building a practice-instance generator that lets the model learn and lets us test fast.
What we learned That machine learning works best here as a fast warm-start rather than the whole solution, and that a strong local evaluation setup — testing our own solutions without waiting on the server — is the biggest driver of fast improvement.
What's next for YardMind — Learning-Guided Shipyard Block Optimizer Building the instance generator and a baseline-beating submission first, then training the placement model and improving the refinement operators throughout the preliminary round.
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for YardMind
