Using Artificial Neural Networks to Study Dependencies in Real-Time Audio and Video Conferencing

Real-time research of modern audio and video conferencing in operational conditions is important for revealing bottlenecks in their functioning. On this basis, recommendations can be made to improve performance indicators, such as productivity, reliability, mode of service, etc. In the study, the architecture of the artificial neural network (ANN) of the type with forward feed and backpropagation of errors was chosen. These networks are made up of several layers of neurons: an input layer, several hidden layers, and an output layer. The study used the Levenberg-Marquardt learning algorithm, which combines gradient descent methods and the Gaussian and Newtonian algorithm. The relationship between "active audio streams - active users" with two neural networks was investigated: The first number of neurons in the hidden layer is 5; The second number of neurons in the hidden layer is 6. The activating function that is used for the neurons of the hidden layer is the tangent hyperbolic, and for the output neuron – the linear function. The neural network with six neurons in the hidden layer in the scope of the training sample, the prediction of ANN is very close to the cloud of points of the original data. It can also be seen that in proximity (about 10% of the training sample range) ANN also has some extrapolation ability, reproducing the behavior of the original data.