On Neural Network Identification for Low-Speed Ship Maneuvering Model

The paper empirically compares two training losses (acceleration- vs state-evaluating) and two RNN-style architectures (standard vs finite-memory window), and reports that (i) training on state rollouts performs better than training on noisy acceleration labels, (ii) the finite-memory variant outperforms the standard RNN on their datasets, and (iii) including randomized maneuvers improves berthing prediction. These findings are supported by experiments and qualitative reasoning, but the paper offers no formal proofs or general guarantees beyond its datasets. The model’s solution provides plausible theoretical sketches: a Lipschitz/Grönwall link from acceleration error to LMSE, OLS-style variance inflation under acceleration-label noise, counterexamples for a universal “no-worse-than” claim, and a Gram-matrix Loewner ordering argument for randomized designs. However, it relies on strong, unstated assumptions (fixed/linearized features, well-specified models, full-rank information matrices, stability, and specific noise structures) and does not fully justify that learned hidden features behave as assumed across the two training objectives. Hence, the paper is incomplete (empirical only) and the model’s proof sketches are insightful but not fully rigorous.