Optimal Transport-based Permutation-Invariant Bayesian Optimization of Offshore Wind Farm Layouts

Technical Summary: Optimal Transport-based Permutation-Invariant Bayesian Optimization of Offshore Wind Farm Layouts

1. Problem Statement

The paper addresses the global optimization of black-box, expensive-to-evaluate, non-convex objective functions defined over layouts. A layout is defined as a point-cloud (a set of $m$ points in a $d$-dimensional space) where the order of the points does not affect the objective function value. This setting is distinct from standard point-cloud optimization where point order matters.

The core challenge identified is the "replication mechanism." In standard vectorization approaches, where a point-cloud $P \in \mathbb{R}^{m \times d}$ is treated as a vector in $\mathbb{R}^{md}$, the objective function $f(P)$ becomes invariant to permutations of the rows of $P$. Consequently, if the true problem has a single global optimum layout, the vectorized formulation creates $m!$ (factorial) global optima. This replication infests the search space with redundant local and global optima, creating an extremely "wiggling" function landscape. This condition hinders the convergence of standard Bayesian Optimization (BO) algorithms, which struggle to distinguish the true underlying function from its permutations.

The authors apply this problem formulation to a real-world industrial case: the optimization of offshore wind farm layouts. Given identical wind turbines, swapping any pair of turbines does not change the Annual Energy Production (AEP), making the problem permutation-invariant.

2. Methodology: PIBO

The authors propose PIBO (Permutation-Invariant Bayesian Optimization), a framework that leverages Optimal Transport (OT) theory to enforce permutation invariance directly within the BO loop.

Key Technical Components:

• Reference Point-Cloud: A reference point-cloud $\bar{P}$ is sampled from a multivariate normal distribution outside the physical search space. This ensures that the cost matrix between $\bar{P}$ and any candidate layout $P$ is in "general position," facilitating unique transport plans.
• Optimal Flow Representation: Instead of optimizing over the point-clouds $P$ directly, PIBO optimizes over the optimal flow $X$. For a candidate layout $P$, the flow is defined as $X = P - \bar{P}$, derived from the optimal transport map $T^*$ that minimizes the squared Euclidean distance between $\bar{P}$ and $P$.
  • Mathematically, the problem is framed using the Monge formulation (an assignment problem) or its linear relaxation, the Kantorovich formulation.
  • The flow $X^*$ is calculated as $X^* = (T^{*\top} - I)\bar{P}$, where $T^*$ is the optimal transport matrix.
• BO over Flows: The Gaussian Process (GP) surrogate model is fitted on the space of optimal flows $X$ rather than the raw layouts. The acquisition function (specifically GP-UCB) is optimized over this flow space.
• Constraint Handling: Physical constraints (e.g., minimum distance $\rho$ between turbines) are reformulated in terms of the flows. The authors prove that if the reference points are sufficiently close, the constraint $\|p_i - p_j\| \geq \rho$ can be rewritten as $\|x_i - x_j\| \geq \| \bar{p}_i - \bar{p}_j \| + \rho$, allowing constraints to be handled directly during acquisition function optimization.
• Iteration: In each iteration, a new flow $X'$ is proposed by the acquisition function. A Kantorovich problem is solved between $\bar{P}$ and the implied layout $\bar{P} + X'$ to ensure the resulting flow is a valid optimal transport plan before evaluating the objective function.

3. Key Contributions

1. PIBO Algorithm: The proposal of a permutation-invariant BO approach based on Optimal Transport theory, specifically designed for optimization over layouts in continuous spaces.
2. Real-World Application: The evaluation of PIBO on the optimization of offshore wind farm layouts, a problem characterized by identical assets and permutation invariance.
3. Empirical Validation: Demonstration that PIBO outperforms vanilla BO (both over physical space and flow space) and baseline methods (Random Search, TPE) in terms of both the optimal objective value and computational efficiency.
4. Open Source: The release of the code to ensure replicability.

4. Experimental Results

The authors evaluated PIBO against several baselines on a test case involving $m=5$ identical wind turbines in a $[0,1]^2$ domain with a minimum separation constraint.

• Performance:

  • PIBO (using an Exponential kernel and uncertainty bonus $\xi=6$) achieved the highest average objective value (AEP) among all methods.
  • Vanilla BO on flows (without permutation invariance enforcement) performed better than Vanilla BO on point-clouds but worse than PIBO.
  • TPE (Tree-Parzen Estimator) and random search methods (LHS, URS) were outperformed by all BO variants.
  • Specifically, PIBO with the Exponential kernel and $\xi=6$ yielded an average best value of 70.77, compared to 70.66 for vanilla BO on flows and 69.82 for vanilla BO on point-clouds.

• Computational Efficiency:

  • Despite the overhead of solving a linear programming problem (Kantorovich formulation) at every iteration, PIBO was approximately twice as fast as the other BO variants.
  • Average time per iteration (GP fitting + acquisition): 347.9 seconds for PIBO vs. ~692 seconds for vanilla BO on flows.
  • The authors attribute this speedup to the simplification of the GP fitting process; by modeling the smooth underlying function over layouts rather than the "wiggling" replicated function over point-clouds, the surrogate model converges more efficiently.

• Qualitative Analysis:

  • Visualizations of the best layouts show that PIBO produces solutions with non-crossing flows (consistent with OT theory), whereas vanilla BO on flows often produced crossing flows, indicating a lack of true permutation invariance.
  • PIBO was the only method capable of inferring a non-linear organization of the physical space into influence regions associated with specific turbines.

5. Significance and Claims

The paper claims that the replication mechanism in standard BO formulations for layout problems is a primary barrier to convergence. By utilizing Optimal Transport to map the search space to a space of optimal flows, PIBO effectively removes this redundancy.

The authors state that permutation invariance is the "key to achieving the best performance" against both standard BO and non-surrogate baselines. Furthermore, they highlight a counter-intuitive finding: enforcing permutation invariance via OT, which adds a linear programming step, actually reduces total computational cost because it drastically simplifies the GP fitting step.

The paper concludes by noting limitations regarding scalability for very large point-clouds ($m > 1000$), where the cost of solving the Kantorovich problem might become prohibitive, suggesting that prefixed sorting could be an alternative in such cases. Future work is suggested for batch-BO acquisition optimization and multi-objective extensions.