Abstract
In this letter, we propose a power allocation scheme for relayed orthogonal frequency division multiplexing with index modulation (OFDM-IM) systems. The proposed power allocation scheme relies on an artificial neural network (ANN) and deep learning to allocate transmit power among various subcarriers at the source and relay nodes. The objective is to minimize the overall transmit power under a set of constraints. We assume all subcarriers at source and relay nodes are independently distributed with different statistical distribution parameters, employing a fixed-gain amplify-and-forward (FG AF) relaying protocol. The adaptive moment estimation method (Adam) is used for back-propagation learning. Analytical and simulation results demonstrate that the proposed scheme provides performance comparable to the optimal solution with significantly lower complexity.
Introduction
Next-generation wireless communication systems, such as 5G and beyond, demand high energy efficiency and coverage. Relayed OFDM with Index Modulation (OFDM-IM) has emerged as a promising technique, utilizing the indices of active subcarriers to convey additional information, thereby enhancing spectral and power efficiency. However, achieving optimal energy efficiency requires precise power allocation across active subcarriers at both the source and relay nodes.
Traditional power allocation methods face significant challenges:
- Heterogeneity: Real-world subcarriers experience varying noise levels and fading characteristics. Most existing models simplify this, but this research addresses the complexity of independent statistical distributions for each subcarrier.
- Complexity: The optimization problem for minimizing power under outage probability constraints is non-convex and non-linear. Solving this in real-time via exhaustive search is computationally prohibitive for dynamic wireless channels.
Methodology
To overcome these challenges, this work treats power allocation as a supervised learning problem, utilizing a multi-layer perceptron (MLP) to approximate the optimal power allocation function.
System Model and ANN Architecture
The system considers a three-node setup (Source $\to$ Relay $\to$ Destination) using a Fixed-Gain Amplify-and-Forward (FG AF) relay. The designed ANN takes statistical Channel State Information (CSI) and average noise power for all active subcarriers as inputs. The output layer provides the optimal power allocation ratios for the source and relay.
The network employs the Sigmoid activation function to map outputs to a [0, 1] range, representing power ratios. Training is conducted using the Adam optimizer to minimize the Mean Squared Error (MSE) between the ANN’s predictions and "ground truth" labels generated via offline exhaustive search.
Dataset Generation
A synthetic dataset was generated based on standard Rayleigh fading channel models. The training set comprised 6,561 samples, while a validation set of 1,000 independent samples was used to evaluate generalizability. The simulation parameters were fixed at $N=4$ subcarriers with $T=2$ active subcarriers.
Results and Discussion
The proposed ANN-based scheme demonstrates a significant reduction in real-time computational complexity, moving from $O(1/\delta^2)$ for exhaustive search to $O(\sum \rho_r \rho_{r+1})$ for the feedforward pass.
Performance and Accuracy
The ANN successfully learned the complex mapping required for power allocation. As shown in Figure 2, the relative error ($\mathcal{E}$) decreases steadily across epochs. The model achieves a relative error of approximately 7.5% when using 128 neurons per layer, which is considered acceptable for practical real-time applications.
Structural Insights
Analysis of the network structure revealed that while increasing the number of neurons improves final accuracy, the number of hidden layers has a more nuanced impact. Fewer layers can sometimes lead to faster initial convergence, though deeper models eventually provide superior precision. Furthermore, the model proved robust across different outage thresholds ($\Psi_{th}$), with higher thresholds facilitating faster convergence during the training phase.
Conclusion and Future Work
The ANN-assisted power allocation scheme offers a near-optimal solution for relayed OFDM-IM systems with significantly reduced operational complexity. While the current approach is limited by the "label bottleneck"—where generating training data for large $N$ is computationally expensive—it provides a strong foundation for real-time energy optimization in small-scale subcarrier blocks. Future research could extend this model to include direct source-destination links and explore modular architectures to handle dynamic changes in the number of active subcarriers.
Summary updated