ANN PREDICTION EARLY STOPPING PROBLEM LEADING TO POOR PEFORMANCE.
4 views (last 30 days)
I am prediction the torque output at propellor shaft with 3 inputs. I have 2 data sets.
I can train both the data sets seperately with a good R squared value, but they perform poorly when tested on the other set.
To solve this I started training on first data set with a 90 % training data from first set and a 10% Validation set ,10% test set from the seond data set.
This gives me a good prediction accuracy but upto a certain limit, Rsquared= 0.75 (between the test, training and validation set).
But as i add more layers or include more neurons, the training stops due to early stopping. Which results in poor performance on all the datasets.
I am unable to improve accuracy.
Vineet Joshi on 26 Oct 2021
Easy stopping usually happens when the models performance is not increasing despite continous backpropagation steps.
Keeping this in mind, there are two ways you can workaround your problem.
Firstly you can change the network architecture and make it such that the model is able to continously improve as training progresses.
Second you can change the criterion of early stopping and make a criterion suitable for your use case.
Kindly refer to the following MATLAB answer link for more information on this.
Hope this helps.