Neural Spatial Interaction Models: Network Training, Model Complexity and Generalization Performance

摘要: Spatial interaction models approximate mean frequencies between origin and destination locations by using origin-specific, destination-specific spatial separation information. The focus is on that are based the theory of feedforward neural networks. This contribution considers functional form models, including specification activation functions, discusses problem network training within a maximum likelihood framework involves solution non-linear optimization problem. requires evaluation log-likelihood function with respect to parameters. Overfitting likely occur in models. To avoid this recommends controlling model complexity either regularization or early stopping training. A bootstrapping pairs approach replacement may be adopted evaluate generalization performance