Neural Computation
A practical Bayesian framework for backpropagation networks
Neural Computation
| Hi-index | 0.00 |
With the 413 soil liquefaction records with cone penetration testing values collected after strong earthquakes, the Bayesian Regularization Back Propagation Neural Networks (BRBPNN) method was presented to evaluate the soil liquefaction potential in this paper. Cone resistance (qc), equivalent dynamic shear stress (τ / σ′0), mean grain size (D50), total stress (σ0), the effective stress (σ′0), earthquake magnitude (M) and the normalized acceleration horizontal at ground surface (a / g) are used as input parameters for networks. Four networks are constructed for different source of input data. The model M7 seems more efficient for the given data, since it only contain 109 records. The model M5 contains 413 samples, and the correct ratio for training data and testing data are 88.5% and 90% respectively. By compared with the square of the weight of the input layer for each network, the importance order of the input parameters should be qc,M,σ′0,σ0,a / g,τ / σ′0 and D50.