Neural architecture search-based BP neural network for underwater glider motion parameter generation

This paper presents a new approach for generating energy-efficient underwater glider motion parameters. First, a valid dataset is built by extracting key input features—path segments, ocean currents, and seafloor topography—while using pitch angle and diving depth as output variables. We then propose a neural architecture search-based back propagation neural network (NAS-BPNN) to enhance the model’s learning and generalization capabilities. In this approach, the number of hidden layers and neurons, activation functions, and training algorithms are treated as decision variables. Encoding-decoding strategies and a fitness evaluation method are designed for these variables. The success history-based adaptive differential evolution algorithm is applied to optimize the network architecture. Simulation experiments demonstrate the effectiveness of the dataset and verify the importance of neural architecture search. Furthermore, NAS-BPNN outperforms the two latest BPNN methods, achieving the highest determination coefficients for motion parameters.