Master thesis: Design and Analysis of Multiplicative Non-orthogonal Multiple Access Schemes

All code concerning my master thesis at CEL:

Implemented is a Neural network training a multiplicative modulation format:

received = (mod1 + n_1)*mod2 + n_2

Init

Before you start, the figure folder has to be created. Run:

cd Multiple_NOMA/Multiple_NOMA
mkdir figures

If you are planning on running the code on cpu, run:

pip install -r requirements.txt

If you want to run on gpu, cupy is installed additionally. Run:

pip install -r requirements_gpu.txt

Additionally, install pytorch according to their website.

Overview

Training and plotting can be done by running split_test.py. We recommend running this file to get accostumed to the code.

Main Components are:

• Plotting of the Fourier transform for different pulseshaping functions: sinc, rect, gauss

• Folder Multipl_NOMA:

  • training_routine has function Multipl_NOMA which is described below

  • training_routine_additive has an equivalent function but uses conventional NOMA, system is: received = (mod1+n_1)+(mod2+n_2)

  • theoreticalMNOMA has the training of a decoder for the theoretical design proposal for a "16 QAM"

  • waverform_training_routine enables training with pulseshaping functions and chromatic dispersion

  • canc_compare: in this script, the different cancellation methods are run successively and are compared

  • design_eval: compares free learning with the design proposal

  • noise_weight, noise_sweep: Noise evaluation of the system

  • other files have content which is used in Mulipl_NOMA etc.

Multipl_NOMA(M=4,sigma_n=0.1,train_params=[50,300,0.005],canc_method='none', modradius=1, plotting=True, encoder=None):

M

is a tensor with the number of symbols of each user: example M=torch.tensor([4,4])

sigma_n

this tensor defines the noise

train_params=[num_epochs,batches_per_epoch, learn_rate]

The training parameters fix the batch sizes, number of training epochs and the learn rate

canc_method

canc_method is the chosen cancellation method:

• division cancellation: canc_method='div'
• no cancellation: canc_method='none'
• cancellation with neural network: canc_method='nn'

modradius

modradius is the permitted signal amplitude for each encoder If the design proposal from thesis is not injected, chose modradius=torch.ones(len(M))

M, sigma_n, modradius are lists of the same size

plotting=False

This enables/disables the creation of the plots:

• Generalized Mutual Information (GMI) during training
• Symbol error rates (SERs) of different users during training
• resulting constellation diagram for best implementation
• constelation diagrams of different users (constellation_base) for best implementation
• decoder decision regions

Plots are saved in figures folder in PDF format

encoder

trained encoders can be injected in the system. The learning rate will be reduces by a factor of 0.01 for the injected encoders, as it is assumed that they are already trained. Injected encoders are added to the end of the list of encoders: They are assumed to be closer to the receiver than new encoders.

example: encoder=[enc1, enc2]

NN training

loss function

A bitwise loss-function and a symbol-wise loss function were implemented. All simulations use the bitwise loss function in order to optimize the bitmapping The loss function is customloss.