Structure Design of Adaptive Probabilistic Modular Neural Network

To address the challenge of online modular neural networks being unable to efficiently handle high-dimensional data in real-time, this paper proposes an Adaptive Probability Modular Neural Network (APMNN). Firstly, the parameters of each Gaussian component within the model are updated in real-time based on probability density, enhancing the network's ability to adapt to time-varying data and improving clustering quality, thereby achieving online partitioning of the sample space. Secondly, through probability membership sets, condition data of different components are assigned to the corresponding subnetworks for learning. Finally, ensemble weights are used to integrate the outputs of different subnetworks to obtain the final prediction results. Predictions on artificial datasets demonstrate that this network can effectively partition high-dimensional data in real-time, achieving higher accuracy and a more streamlined structure compared to traditional modular neural networks.