Akademik Veri Yönetim Sistemi
IEEE Access, 2026 (SCI-Expanded, Scopus)
Underwater acoustic positioning is challenging because environmental variability and multipath propagation can induce heavy-tailed localization errors. This paper proposes a physics-based, simulation-driven feed forward multilayer perceptron framework that treats sound-speed variability and multipath not as disturbances to be suppressed, but as informative components of the mapping from multi-receiver time-of-arrival (TOA) measurements to 2D range–depth coordinates. Using BELLHOP-generated TOA datasets, 108 neural configurations spanning hidden layer depth, neuron allocations, activation patterns, and three training algorithms were benchmarked. A trade-off analysis between training cost and localization accuracy, complemented by tail-robustness assessment, revealed clear optimizer-dependent behavior: Bayesian regularization provided the most favorable overall balance between typical accuracy and robustness, Levenberg–Marquardt remained competitive in several configurations but showed less uniform gains across architectural variations, and scaled conjugate gradient offered markedly shorter training times while operating in a substantially higher error regime. These results show that deployment-oriented model selection should consider not only median localization accuracy, but also tail behavior. Under the examined conditions, the best configuration, obtained with Bayesian regularization, achieved a median 2D positioning error of approximately 6.2 m with a p95 of approximately 10.4 m, representing a 36.1% reduction in median error and a 33.8% reduction in p95 compared to the best Levenberg–Marquardt model (9.7 m median, 15.7 m p95). In contrast, the best scaled conjugate gradient configuration remained above 300 m median 2D error. Overall, Bayesian regularization emerged as the most reliable optimizer within the evaluated configuration set.