A New Collaborative Neural Network: Using HDMR for Dynamic Structural Adaptation*

In modern engineering and scientific research, accurate approximation of high-dimensional nonlinear models is crucial for understanding complex systems. As the model dimension increases, traditional approximation methods encounter the “curse of dimensionality”. To address this issue, we propose a collaborative neural network model based on the High Dimensional Model Representation (HDMR) method. In this model, each hidden layer in the feedforward neural network is dedicated to fitting individual order component functions of HDMR. During the training process, ineffective submodules will be dynamically pruned, while retaining submodules that significantly impact the model output. In addition, the proposed model structure can be further improved by integrating multiple base models and conducting sensitivity analysis. Finally, we compare the proposed model with some traditional HDMR algorithms on multiple test functions, and the results show that our model demonstrates higher computational efficiency and accuracy.