Combinatorial optimization of input features and learning parameters for decorrelated neural network ensemble-based soft measuring model

摘要

The decorrelated neural network ensemble (DNNE) algorithm can be used to construct an effective soft measuring model through an analytical solution comprising submodels of multiple randomized neural networks. However, DNNE exhibits one major shortcoming: the scope of the random input weights and biases is set to a default range of [-1, 1], which cannot ensure the universal approximation capability of the resulting prediction model. The three other learning parameters including the number of the ensemble submodels' hidden nodes, ensemble size, and regularizing factor of DNNE are also data dependent. Moreover, DNNE suffers from high computation complexity when processing a high-dimensional large-scale dataset. Feature selection can improve the interpretation of the model. To address the above-mentioned problems, a combinatorial optimization method based on the adaptive genetic algorithm is used to simultaneously optimize the input features and learning parameters of the DNNE model. The evolution processes of these modeling parameters are demonstrated in detail. Simulation results based on four datasets with different dimensions and sizes validate the effectiveness of the proposed approach. The results also indicate that the random parameter scope assignment significantly influences the generalization performance and the other learning parameters of the prediction model. (c) 2017 Elsevier B.V. All rights reserved.