Hi all, i am doing a project that models the removal efficiency of wastewater with different operation conditions. I am using Rstudio with keras packages for the modeling. I have some question pertaining to the model creation. I have 5 independent variable (operation conditions) and 1 dependent variable.
1) what is a ball park figure for the number of nodes/ layers. (i understand that there is no right/ wrong answers to this question. i have read some suggestions from stack overflow on the ballpark figures to Xno. of layers/ Yno. of variables. but with no explanation or statistical explanation for it)
2) I also understand that there are means to trimming the no. of nodes that has low activation rate. how am i able to view it via the keras model?
3) is there any means to make my model more accurate?
Please let me know. Thank you very much in advance.
Related to your query I would like to suggest you some links which will be useful for solutions to your tasks: https://www.mathworks.com/help/deeplearning/examples/train-a-convolutional-neural-network-for-regression.html;jsessionid=2c56a20c391fba5cd00177e8cd34
https://www.mathworks.com/products/deep-learning.html
ANN can be used with regularization to prevent over fitting.
Divide data into 3 parts
train, test, validate (use only once after finalizing the model)
You can try different node sizes and use a performance criteria like to select the best node size
Hi Muhammad Ali thank you for your reply and the links provided.
Chukwudi Obite Sohan Singh Rajpurohit thank you for your reply and your proposed solutions. yes adjustments to the node size can increase performance however might potential lead to the problem of overfitting.
my issue is that i would like to know if there is a guideline (ie 4 variable) the layers the no. of nodes conventionally should lie within what ranges. this will let me have a better judgement if i have a situation of under/overfitting. Yes ANN modeling there is no one answer fits all solution, however i am just curious that is the general sensing to the no. of nodes per variable input. Do let me know if you guys know anything about this. Thank you
there are some empirical formulas to determine number of weights and hidden layers. But before applying the ANN, try to understand the nature of data, collinearity in the dataset, correlation of input with output variables.
Yet there are some empirical equations to determine number of hidden nodes and layers based on number of observations and number of variables.
also the selection of activation function matters a lot
One important thing to mention related to the accuracy is
don't touch the final set of validation data
validation data can be used once
otherwise your training model will also memorize the validation data.
make the model with training data
refine with test data
but use the validation set only once.
Brandon Lee I do not know of any guideline but I know you ca control over fitting by introducing a decay parameter to the error function you are minimizing to estimate the weights of the network.
- Badges
- Science topic