I am working on predicting concrete mechanical properties through ANN. I am using MatLAB to program the ANN. How to improve the accuracy of the model, I tried different combinations of hidden layers, used different activation functions, but unfortunately the RMSE for training and validation is too high. I am looking for a technique or approach which can provide me the optimum results or help me find the optimum parameters.
Dear Fatima Khalid,
1. Try to increase the number of data used for ANN (input and output). You can use curve-fitting toolbox in MATLAB to fit a better curve (use it instead of ANN for prediction or) and then use that curve to generate more data.
2. Another way is to "tune" your input and output data, i.e. Plot a scatter graph of your data, check if any data is too far from a mean position, if so, remove them.
Hi Fatima,
I agree with Siddharth Ramachandran answer that increasing the number of data should help improve model efficiency. A few points based on my experience:
1) Check if your data is normalized (pre-processing) before training the model. If not, try normalizing your data first.
2) Try to use 70-80% of the data for training.
3) Try using different algorithms to optimize your ANN model. (In MATLAB, nntoolbox I think uses LM method by default)
Try WEKA. It fives you the flexibility of different algorithms and easy of modifying various modeling parameters.
I agree with the previous answers. I will also suggest that you normalise the data to ensure that the features range from 0 to 1. This will drastically minimise the mean squared error for training, testing, and validation.
- Badges
- Science topic