Machine learning centered energy optimization in mobile edge computing: a review

Abstract

Current literature reviews on machine learning-based approaches for mobile edge computing (MEC) energy optimization often lack in-depth gap analysis and fail to identify trends or offer actionable insights. Most focus narrowly on comparing MEC frameworks without critically evaluating or benchmarking prior research. This review contributes by addressings these gaps via analysis of existing reviews and related studies, with a focus on ML models, research objectives, evaluation metrics, datasets, tools, and gap identification. The review method follows a systematic literature review (SLR) using the PRISMA framework for transparency and reproducibility. Key findings reveal persistent challenges in energy consumption, computational overhead, cost, and poor performance in accuracy, QoS, latency, scalability, and carbon footprint. Deep reinforcement learning (DRL) emerges as the most commonly used model (55%), while TensorFlow (35%) is the most adopted tool, valued for its flexibility and robust community support. The AudioSet dataset is frequently used (28%) due to its compatibility. However, methodology limitations include dependency on study quality and exclusion of grey literature, context sensitivity. The review concludes by recommending advanced solutions such as serverless computing, liquid cooling, containerization, software-defined power, quantum computing, and blockchain to drive future MEC energy optimization.