Dynamic and Energy-Efficient Management of Programmable Network Infrastructure Utilizing Reinforcement Learning Techniques

Abstract

Network Splitting is a critical concept in modern telecommunications, particularly within the context of 5G and beyond, where it enables the decomposition of the Radio Access Network (RAN) into multiple virtualized subunits, each with distinct and heterogeneous requirements. This approach, especially under the disaggregated O-RAN (Open Radio Access Network) architecture, ensures the efficient operation of diverse services in parallel. The virtualization of the network allows for dynamic splitting of the functions of a slice between the network nodes, enabling the integration of advanced optimization mechanisms to manage resources effectively, maximize performance, and control power consumption. In this context, the present thesis proposes a solution based on Reinforcement Learning techniques. The developed agent is responsible for the dynamic admission control and placement of network slices, as well as the optimal splitting of Virtual Network Functions (VNFs) of each slice across the nodes of the O-RAN architecture. The proposed model is trained to make decisions with the aim of minimizing energy consumption and optimizing resource utilization, adapting to varying patterns of slice request generation. Experimental results demonstrate the agent’s ability to make rational and optimal decisions, tailored to the given environmental conditions. Additionally, the importance of the dynamic approach is confirmed through comparison with a static agent, which underperforms in the evaluated performance metrics.