Self-Renewal Machine Learning Approach for Fast Wireless Network Optimization

The throughput maximization in multi-hop wireless networks is largely limited by interference due to the reuse of the channel resources. Although machine learning (ML) can accelerate the optimization of wireless network capacity, the existing system can become limited because of insufficient knowledge from available data. We propose a self-renewal ML (SRML) method that incrementally improves the throughput maximization of future optimization instances through the design of a data selection algorithm for scheduling structure classification and application identification model retraining. With one round of implementation, the SRML method outperforms the fixed ML (FML) method, random (RAND) method and Greedy Heuristic method in the multi-commodity flow deployment setting with an average achievable throughput of 100% for small flows and at least 79% for large flows, relative to the delayed column generation (DCG) benchmark algorithm, while reducing the computational complexity and achieving a high solution efficiency. By leveraging the transfer learning of parameters during self-renewal, the computational cost of model training is reduced by at least 71.09%.