level: technical
bayesian optimization is a common tool for tuning designs when evaluating each option is slow or expensive. but standard bayesian optimization ignores symmetries in the problem. for example, when placing identical wind turbines in a farm, swapping any two turbines does not change the energy output. this symmetry means many different layouts are effectively the same, yet the optimizer wastes time exploring them separately.
the work introduces a permutation-invariant bayesian optimization approach that uses optimal transport to compare layouts. instead of treating turbine positions as an ordered list, it represents a layout as a probability distribution over the site. the distance between two layouts is then measured by the wasserstein distance, which naturally ignores the order of turbines. this distance is plugged into a gaussian process surrogate model, letting the optimizer focus on truly different configurations.
tests on realistic offshore wind farm scenarios show the method finds better layouts with fewer evaluations than standard bayesian optimization. it also outperforms other symmetry-aware baselines. the approach is not limited to wind farms; it can apply to any layout optimization where the order of points does not matter, such as sensor placement or facility location. the code is available to encourage further use.
why it matters: handling symmetry correctly can cut the number of expensive simulations needed to find good designs, saving time and compute in industrial optimization tasks.