MONTE CARLO SIMULATIONS AND FACTOR ANALYSIS TO OPTIMIZE NEURAL NETWORK INPUT SELECTIONS AND ARCHITECTURES

摘要

Artificial Neural Networks (ANN) can model and predict nonlinear coastal processes with equal or better skill than other techniques such as multilinear regression. ANN structural optimization is a non trivial problem. The challenge is to find and demonstrate that the selected structure of ANN is 'good enough'. We employ Monte Carlo simulations of ANN to optimize the input selection and the number of hidden neurons. ANN models were used to forecast water levels for stations along the Texas coast based on previous hourly water levels and winds. First, we randomly simulated 1000 neural nets with different numbers and types of inputs and with different numbers of neurons in the hidden layer. After training of each neural net on one year of hourly data, we tested its performance on seven other years of data. This yielded data about the quality of the predictions made by each ANN using NOAA criteria: the RMSE and of the Central Frequency. As a next step we used factor analysis to explore the impact of different ANN designs on the quality of predictions.