I am building a Neural Network Model for approximating SINR value at a specified location in 2D. The scenario consists of a central gNB placed at (1500, 1500), and other 6 gNBs are placed surrounding it in a hexagonal manner with ISD equal to 1000 m. The UEs are placed around the central gNB at distances from 30m to 500m in steps of 20m, along X and Y. I thus have 2500 points at which I am getting SINR.
For training, X and Y coordinates of UE and their values obtained from NetSim simulation have been used. Thus, the features are coordinates (X, Y), while the Label is SINR.
The NN model architectures I have tried using (i) 2-5 layers (ii) Activation functions: ReLU, LeakyReLU, tanh, sigmoid
(iii) Different units size: 32 to 128, (iv) Dropout at various layers from 0.4 to 0.5,
However, the minimum MSE obtained was around 37 while training, which is very high.
Here are some ways to reduce the MSE and improve the R2 measure:
Maitri Saraf You might attempt the following to lower the MSE and enhance the R2 measure:
1. Use additional data: You might try gathering more data points or training the model with more data. This may aid the model in better learning the link between the characteristics and the target variable.
2. Feature engineering: You might try to create new features that are more relevant to your problem by combining current ones. This may aid the model in better learning the link between the characteristics and the target variable.
3. Experiment with alternative model architectures: You can use a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), or a combination of both.
4. Fine-tune the hyperparameters: You may experiment with changing the learning rate, number of hidden layers, and number of neurons in each layer.
5. Regularization: To avoid overfitting, try using regularization techniques such as L1 or L2 regularization.
6. Cross-validation: To acquire a better estimate of how well your model will perform on unseen data, consider employing cross-validation approaches.
7. Early stopping: To minimize overfitting and enhance model performance, attempt early halting.
8. Ensemble methods: To boost overall performance, you might try mixing numerous models.
It is crucial to remember that finding the ideal model and hyperparameters is an iterative process, and you may need to test several combinations before you discover the one that works best for you.
I think, 2500 data points may not be sufficient to train a NN. You can get more measurements from NetSim by
a) perturbing the UE locations randomly to get new X, Y vs. SINR
b) changing the random seed if you have any randomness (like fading) in your scenario
Thank you so much for your inputs.
I could edit the model architecture suitably.
The measure I could obtain on test data obtained from split:
- MSE around 0.2
- R2 around 0.98 to 0.99
Next, I want to know how to train the DNN and "save" it. This way I can "deploy" the trained DNN, and not train the DNN every time.
I want to pass data from a c file to the trained model and then obtain the output from the c file.
You can save the trained model in Keras using the model.save method. The saved model can then be loaded later and used for prediction without the need for re-training.
It will be saved as a h5 file. The saved model file contains the architecture of the model, the weights, and the optimizer's state. You can then load the saved model and use it for prediction.
This link explains the NetSim C to Python interface.
https://support.tetcos.com/support/solutions/articles/14000131736-can-you-suggest-resources-for-netsim-python-interfacing-
It can be a starting point.
- Badges
- Science topic
- Similar topics
- Correction