Influence of Neural Network Learning Algorithms on High Power Amplifier (HPA) Predistortion Performance

In this article, we propose a feedforward neural network model designed to approximate the inverse transfer characteristic of a High-Power Amplifier (HPA) in order to linearize it using Digital Predistortion (DPD). This approach is particularly relevant for next-generation communication systems, such as those employing OTFS (Orthogonal Time Frequency Space) modulation envisioned for 6G, whose signals exhibit large amplitude variations that exacerbate amplifier nonlinearities. The performance of predistortion heavily depends on the learning algorithm used to train the neural model. We compared three optimization algorithms: Gradient Descent, Gauss-Newton, and Levenberg-Marquardt. The amplifier is modeled using the Rapp model. The neural network architecture consists of a single input neuron, a hidden layer with ten neurons using the hyperbolic tangent activation function, and a linear output neuron. Training and simulations were carried out in MATLAB, and the performance of each algorithm was evaluated using the Mean Squared Error (MSE) criterion, which quantifies the deviation between the ideal transfer characteristic of a linear amplifier and the characteristic obtained after predistortion. The results clearly show that the Levenberg-Marquardt algorithm provides the best approximation of the predistortion function, achieving an MSE on the order of 4.2708×10⁻⁸, significantly outperforming Gauss-Newton 1.0481×10⁻⁴ and Gradient Descent (0.0272). This superior performance is attributed to Levenberg-Marquardt’s ability to combine the robustness of Gradient Descent with the fast convergence of Gauss-Newton, while avoiding local minimum and issues related to poor synaptic weight initialization.