Jenkins William F, Gerstoft Peter, Park Yongsung
Scripps Institution of Oceanography, University of California San Diego, La Jolla, California 92093, USA.
J Acoust Soc Am. 2024 Aug 1;156(2):812-822. doi: 10.1121/10.0028177.
Geoacoustic inversion can be a computationally expensive task in high-dimensional parameter spaces, typically requiring thousands of forward model evaluations to estimate the geoacoustic environment. We demonstrate Bayesian optimization (BO), an efficient global optimization method capable of estimating geoacoustic parameters in seven-dimensional space within 100 evaluations instead of thousands. BO iteratively searches parameter space for the global optimum of an objective function, defined in this study as the Bartlett power. Each step consists of fitting a Gaussian process surrogate model to observed data and then choosing a new point to evaluate using a heuristic acquisition function. The ideal acquisition function balances exploration of the parameter space in regions with high uncertainty with exploitation of high-performing regions. Three acquisition functions are evaluated: upper confidence bound, expected improvement (EI), and logarithmically transformed EI. BO is demonstrated for both simulated and experimental data from a shallow-water environment and rapidly estimates optimal parameters while yielding results comparable to differential evolution optimization.
在高维参数空间中，地声反演可能是一项计算成本高昂的任务，通常需要进行数千次正向模型评估才能估计地声环境。我们展示了贝叶斯优化（BO），这是一种高效的全局优化方法，能够在100次评估内而非数千次评估内估计七维空间中的地声参数。BO迭代地在参数空间中搜索目标函数的全局最优值，在本研究中该目标函数定义为巴特利特功率。每一步都包括将高斯过程代理模型拟合到观测数据，然后使用启发式采集函数选择一个新的点进行评估。理想的采集函数在具有高不确定性的区域平衡对参数空间的探索与对高性能区域的利用。评估了三种采集函数：上置信界、预期改进（EI）和对数变换后的EI。针对来自浅水环境的模拟数据和实验数据展示了BO，它能快速估计最优参数，同时产生与差分进化优化相当的结果。
