I have made a predictive model of liquefaction analysis using Artificial Neural Networks. For time series analysis with neural networks, there is a model calibration coefficient like the Nash–Sutcliffe model efficiency coefficient. But my question is, for neural network model fitting, is there any calibration coefficient like the Nash–Sutcliffe model efficiency coefficient ? Please help me by giving suggestions or advice.
Nash-Sutcliffe is basically and observed-expected kind of thing right? If that's what you're looking for you could test the NNet's fit on a separate validation set - if you've modeled the underlying structure of the system the NNet should fit new data (that wasn't used to train it) well, if you're over-fitting noise it should perform poorly.
Assessing the predicted vs actual would be up to you, if I'm only predicting a continuous variable I tend to approximate the empirical success with a simple regression between the predicted and actual data (because the relationship should be perfectly linear with an R^2 of as close to 1 as possible).
You can also use AICc for neural networks (depending on how they're trained), there's an R package called NeuralNet that can give AICc. However AICc for a single model isn't really that helpful.
- Badges
- Science topic
- Similar topics
- Engineering
- Civil Engineering