MFDROO: Matrix Factorization-Based Deep Reinforcement Learning Approach for Stable Online Offloading in Mobile Edge Networks

Data-intensive applications coupled with limited mobile resources make opportunistic computation offloading imperative. Therefore, efficient and reliable offloading strategies are crucial for achieving optimal performance in terms of stabilizing data queues at various data rates, efficient task management and reduced waiting times for users. This paper proposes a novel algorithm, Matrix Factorization-Based Deep Reinforcement Learning Approach for Stable Online Offloading in Mobile Edge Networks (MFDROO), that combines Orthogonal Non-negative Matrix Factorization (ONMF) and a Deep Reinforcement Learning approach (DRL) to address the problem of stable online offloading in large networks. The proposed algorithm utilizes ONMF to model network resource heterogeneity by decomposing it into a set of features, capturing the underlying structure of data traffic flows by removing irrelevant information from data. This reduces computational overhead and improves performance. Additionally, MFDROO incorporates a DRL agent to learn optimal offloading decisions over time. By utilizing ONMF in conjunction with DRL, MFDROO overcomes online offloading challenges. It optimizes user computation rates and enhances system performance. Additionally, it maintains data-queue stability for large networks or higher data rates. Large-scale network simulations were extensively conducted to demonstrate MFDROO effectiveness. Maintaining stability while scaling up the users number is a challenge in network computation. Our results demonstrate that MFDROO tackles this challenge and ensures that even with a 3.3% increase in user numbers (up to 100 users), computation networks remain stable and outperform other existing algorithms in terms of utility maximization and stability. This improvement ensures optimal performance and provides scalability and efficiency for various applications.