Dear all
I have created a custom script for Neural Network for my dataset with 36data as training set and 6data as test set.
My activation functions for both hidden layer and output layer is taken as log-sigmoid function.
I have 2 input nodes and 1 output node. I have reached to an optimum architecture of 2-15-1 with learning rate of 0.6 and momentum constant of 0.55.
After both training and testing, MSE for training was found to be 0.0005 and that for testing was found to be 0.0006.
But later on I have tried to test the network with another random example in which case I got a huge error percentage of 40-60%. I don't know where I am making mistake. Do I need to increase the number of datasets in the training ?
N.B I need to run GA using NN as my objective function.
Please help.
Dear Rajdeep Ghosh,
In my knowledge, the performance of the model will vary from data set to a data set. Anyhow, what is the ratio of training and testing you have considered? You change the ratio to either 75-25 or 85-15 and check it again?
Thanks
It seems to be a case of over-fitting. Have you used a validation set during training? Also instead of using a training/testing set you can use 10-fold cross-validation to check whether over-fitting is occurring.
I would definitely recommend revising your train/test split and experiment with it to find the optimal split. Secondly, try a different activation function.
Dear all....Thank you for you wonderful suggestions. I was actually doing some mistake while normalizing the input variables. Now training set has approx 5% noise readings and both training and testing data shows an R-squared of around 0.82.
But still what do you think if more datasets could have been taken to reduce the noise further which might give poor prediction in GA.
This link talk about:
Add PyTorch/TensorFlow/Keras custom nn modules for various QML templates #166
https://github.com/PennyLaneAI/pennylane/issues/166
- Badges
- Science topic
- Similar topics
- Organizational Psychology
- Training