Multi-Objective Dataset Optimization for Learning-Based DC-Bias Prediction in DCO-OFDM

Technical Summary: Multi-Objective Dataset Optimization for Learning-Based DC-Bias Prediction in DCO-OFDM

Problem Statement
In Visible Light Communication (VLC) systems utilizing Direct Current Optical Orthogonal Frequency Division Multiplexing (DCO-OFDM), the selection of the DC-bias is a critical design parameter. It directly influences interdependent performance metrics: Bit Error Rate (BER), Peak-to-Average Power Ratio (PAPR), and clipping noise. While existing literature has extensively addressed DC-bias optimization using heuristic algorithms or machine learning (ML) for specific transmission instances, a significant gap remains in the dataset level. Current approaches typically rely on exhaustive, brute-force simulations to generate training data, often resulting in redundant samples and a lack of systematic optimization to capture multi-objective trade-offs. Furthermore, existing ML-based works focus on improving prediction models without validating whether the underlying dataset itself is optimally structured for learning. There is a lack of a systematic framework to generate, select, and validate an optimized dataset that reflects the complex relationships between signal statistics and optimal bias values for generalized learning.

Methodology
The authors propose a systematic framework that shifts the focus from instance-level optimization to data-centric modeling. The methodology proceeds through four main stages:

1. Raw Dataset Generation: A high-dimensional raw dataset is constructed via exhaustive simulation in MATLAB. The process involves generating 5,000 DCO-OFDM symbols with 4-QAM modulation and a symbol length of 2048. DC-bias values are swept from 2 dB to 12 dB (in 0.05 dB increments) across Signal-to-Noise Ratio (SNR) values ranging from 8 dB to 25 dB. For each configuration, statistical features (minimum, maximum, mean, and standard deviation of the pre-bias signal) and performance metrics (average BER, PAPR, and clipping noise power) are extracted. This results in a raw dataset of 5000 symbols $\times$ 201 bias levels.

2. Multi-Objective Dataset Optimization: The problem is reformulated as a multi-objective optimization task to select the most suitable samples from the raw dataset. An objective function $J$ is defined as a weighted sum of the performance metrics: $J = 0.5 \times \text{BER} + 0.3 \times \text{PAPR} + 0.2 \times \text{Clipping Noise}$. Six optimization techniques are applied to extract optimal samples (rows) for each symbol:

   • Genetic Algorithm (GA)
   • Particle Swarm Optimization (PSO)
   • Mutual Information (MI) with static weights (MI-static)
   • Mutual Information (MI) with dynamic weights (MI-dynamic)
   • Pareto Optimization
   • Technique for Order Preference by Similarity to Ideal Solution (TOPSIS)

3. Multi-Stage Evaluation Framework: The six resulting datasets are evaluated through a rigorous, multi-stage process:

   • Statistical Analysis: Examining the behavior of the selected optimal bias values across SNR levels to ensure adaptability and physical meaningfulness.
   • Performance-Based Evaluation: Comparing the resulting BER, PAPR, and clipping noise across the different optimizers.
   • Composite Scoring: A unified score derived from normalized metrics to assess overall trade-off balance.
   • Statistical Significance Testing: A non-parametric Wilcoxon signed-rank test is applied to compare the BER distributions of the top-performing methods.

4. Machine Learning Validation: As a final validation criterion, the datasets generated by the top contenders are used to train seven regression models (Linear Regression, Decision Tree, Random Forest, Gradient Boosting, XGBoost, SVR, and kNN). The datasets are evaluated based on their ability to support accurate predictive modeling, measured by $R^2$, RMSE, and MAE.

Key Contributions
The paper outlines five primary contributions:

1. Dataset-Level Reformulation: The work reformulates the DC-bias problem from instance-level optimization to a dataset-level optimization, aiming to construct a structured dataset that captures underlying relationships for better ML generalization.
2. Systematic Data Generation: The creation of a high-dimensional dataset sweeping DC-bias across multiple SNR conditions, extracting both statistical signal features and performance metrics.
3. Comparative Optimization Analysis: A comprehensive comparison of six distinct optimization techniques (GA, PSO, MI-static/dynamic, Pareto, TOPSIS) for extracting optimal samples based on a weighted multi-objective function.
4. Novel Evaluation Framework: The introduction of a multi-stage evaluation strategy that combines statistical analysis, performance metrics, composite scoring, and statistical significance testing, culminating in ML-based validation to assess dataset suitability for predictive modeling.
5. Foundation for Future Work: The establishment of a structured, optimized dataset that serves as a baseline for future data-driven DC-bias estimation and adaptive communication systems.

Results
The evaluation yielded the following findings:

• Optimizer Behavior: GA, PSO, and Pareto-based methods produced nearly constant bias values across SNR levels, indicating a lack of adaptability to channel variations. In contrast, MI-static, MI-dynamic, and TOPSIS demonstrated monotonic increases in bias with SNR, consistent with expected system behavior.
• Performance Trade-offs: While MI-dynamic achieved low PAPR and clipping noise, it resulted in inferior BER performance due to bias overestimation. GA and PSO minimized individual metrics (PAPR/Noise) but failed to balance the trade-off, leading to poor BER.
• Top Contenders: MI-static and TOPSIS emerged as the most effective methods, achieving a balanced trade-off. Statistical significance testing (Wilcoxon signed-rank) revealed that MI-static performed better in low-SNR regimes, while TOPSIS outperformed in medium-to-high SNR regimes.
• ML Validation: When used as training data for regression models, the TOPSIS-based dataset consistently outperformed the MI-static dataset. Across multiple models (e.g., XGBoost), the TOPSIS dataset achieved higher $R^2$ values (approx. 0.9958 vs. 0.9946) and lower error metrics (RMSE and MAE). This indicates that the TOPSIS-selected data possesses superior structural consistency and learnability.

Significance and Claims
The authors claim that this work addresses a critical gap in the literature by explicitly optimizing the dataset rather than just the model or the instance. The significance of the work lies in demonstrating that the quality and structure of the training data are decisive factors in the success of learning-based DC-bias prediction.

The paper concludes that while MI-static and TOPSIS offer comparable communication performance, the TOPSIS-based dataset is the optimal choice for learning-based frameworks. This selection is driven not merely by communication metrics but by the dataset's superior capability to support predictive modeling, evidenced by higher accuracy and lower error rates across diverse regression models. The proposed framework provides a foundation for developing adaptive, data-driven optical OFDM systems, moving beyond brute-force simulation toward intelligent, optimized data construction. The authors note that future work will utilize this selected dataset to develop real-time prediction models for unseen system conditions.