Jul 2026· Signal Processing and Communications Applications Conference· pp. 1-4· 0 citations· 19 references
Abstract
This study introduces an effective optimization technique for identifying the optimal hyperparameters of a deep neural network (DNN) designed to model the behavior of power amplifiers (PAs). Hence, an innovative and efficient yield-analysis-based approach is proposed to improve both the modeling accuracy and the digital predistortion (DPD) performance of PAs. The method leverages a long short-term memory (LSTM) DNN architecture to capture the nonlinear and memory effects inherent in PA systems, thereby enhancing overall performance and linearization capability. To achieve optimal training, multiple optimization strategies are systematically applied to determine the most suitable hyperparameters, such as learning rate, network depth, and neuron configuration. To validate the effectiveness of the proposed method, a PA operating in the 1.8 GHz to 2.2 GHz frequency range is considered, for which extensive simulations and evaluations demonstrate that the optimized DNN achieves minimal modeling error while significantly improving DPD performance. The results confirm that the proposed framework offers a reliable and efficient solution for PA modeling and linearization, outperforming conventional techniques in terms of accuracy and consistency.
Due to the continuous increase in communication bandwidth and the use of highly efficient yet nonlinear power amplifiers, Digital Predistortion (DPD) algorithms are becoming increasingly complex. In particular, neural network (NN) based DPD approaches using Phase-Normalized NN architectures often incur substantially higher computational costs than widely deployed polynomial-based methods, such as the Memory Polynomial (MP) and Generalized Memory Polynomial (GMP) models. To bridge this gap between research performance and practical implementation, we propose a low-complexity Feature Selection NN DPD architecture. The proposed method employs an offline feature-engineering pipeline based on the Least Absolute Shrinkage and Selection Operator (LASSO) and the Minimum Redundancy Maximum Relevance (MRMR) algorithm to construct a compact and informative input representation. Using measured wideband FR3 power amplifier datasets that are publicly released with this work, we demonstrate up to 30% reduction in computational complexity while maintaining comparable linearization performance.
Cel Thys, Rodney Martinez Alonso, A. Alsarraf et al.· 0 citations
In this study, we examine and contrast the effectiveness of different artificial neural network (ANN) topologies for power amplifier (PA) digital pre-distortion (DPD). In particular, we investigate long short-term memory (LSTM) networks, gated recurrent units (GRU), recurrent neural networks (RNN), con-volutional neural networks (CNN), and fully connected neural networks (DNN). For training and assessment, a dataset comprising measured input and output signals from a commercial NXP Doherty PA working in the 3.6–3.8 GHz region with a 16-QAM OFDM signal is utilised. Normalised mean squared error (NMSE), adjacent channel power ratio (ACPR), and model complexity are used to evaluate the models. Simulation results show that while CNNs offer a favorable trade-off between linearization performance and model complexity, GRU and LSTM architectures achieve the best overall NMSE and ACPR improvements, albeit with a higher number of parameters. Power spectral density and AM/AM characteristic analyses further confirm the superior linearization performance achieved using recurrent gated structures.
Hafsa Laakouri, M. Ouadefli, A. Tribak et al.· EPJ Web of Conferences· 0 citations
This work presents a direct-learning architecture of a neural network (NN) digital predistortion (DPD) linearizer for a multiple-input multiple-output (MIMO) system while maintaining low complexity compared to a single-input single-output (SISO) system.
The analysis demonstrates that the RL-based approach, enabled by an effective neural network initialization strategy, surpasses traditional methods and ML-based DPD schemes such as DLA and ILA and provides a scalable and efficient solution for compensating pattern-dependent nonlinearities in high-speed optical communications.
Arash Rabiepoor, L. Rusch, Ming Zeng· IEEE Open Journal of the Com...· 0 citations
Charge-transport models provide quantitative and physically interpretable descriptions of memristive devices but are computationally prohibitive for highly iterative tasks in model-driven design workflows such as parameter space exploration, parameter extraction, and optimization. Here, we investigate recurrent neural networks (RNNs) as efficient sequence-to-sequence surrogates to accelerate dynamic memristive transport models and provide systematic guidance on training strategies, architectural choices, and data requirements. In this context, we focus on long short-term memory (LSTM) and gated recurrent unit (GRU) architectures combined with advanced training and normalization strategies. The results demonstrate that layer normalization substantially improves convergence, training stability, and generalization, whereas chrono initialization degrades performance in this setting. The most robust training behavior is obtained by combining layer normalization with the AdamW optimizer and cosine annealing learning-rate scheduling, with GRU architectures achieving the lowest errors overall. Using the optimized configurations, mean normalized errors below 0.1% are achieved for a five-dimensional parameter space. Accurate performance is retained with limited training data, with fewer than 1000 configurations still yielding mean errors around 0.15%. Increasing the input dimensionality leads to a systematic rise in error from ∼0.1% (5D) to 0.66% (9D), while mean errors remain small, well below 1%. These results establish practical design rules for applying RNN-based surrogates in model-driven design and optimization of memristive devices.
Somayeh Rezaei, Benjamin Spetzler, Patrick Mäder· APL Machine Learning· 0 citations
Digital predistortion (DPD) compensates for nonlinear distortions caused by RF power amplifiers (PAs) for efficient and linear signal transmission. Although generalized memory polynomial (GMP) models are frequently used for DPD, their dimensionality increases with memory depth, which raises computational costs and deteriorates numerical conditioning. Although batch principal component analysis (PCA) reduce this dimensionality, it is unable to adjust to PA characteristics that change over time. We present IPCA-GMP, a framework that combines GMP modelling with incremental PCA (IPCA), where updating mean and covariance estimates recursively and executing rank one eigenspace updates from streaming data. The eigenspace update achieves $\mathcal{O}\left(k D^{2}\right)$ per block cost without a complete eigen decomposition by using a rank one perturbation procedure on the current eigen basis. The results demonstrate that IPCA-GMP compresses the model dimension from $D=224$ to $k=20 (11 \times$ compression), achieving a normalized mean square error of -39.02 dB within 0.41, dB of the full GMP baseline. The condition number is better than $3.3 \times 10^{26}$ to 1,570, and with just 1.8% FLOP overhead compared to the baseline GMP solver, the per-update complexity drops from $\mathcal{O}\left(N D^{2}+D^{3}\right)$ (batch PCA) to $\mathcal{O}\left(D^{2}+k D^{2}\right)$.
Girish Chandra Tripathi, Anindya Saha· International Conference on...· 0 citations