Dynamic Multi-Agent Reinforcement Learning Based Load Balancing on Software Defined Networking

Cloud computing provides on-demand access to computing re-sources via the internet, but efficient resource utilization and traffic man-agement remain challenging due to dynamic workloads. This paper proposes a Dynamic Multi-Agent Reinforcement Learning (MARL)-based Load Balancing (DA-LB) algorithm for Software-Defined Networking (SDN)-enabled cloud data centers. Unlike existing heuristic or single-agent reinforcement learning approaches, DA-LB leverages decentralized cooperation among MARL agents to achieve adaptive load balancing, leveraging SDN's global visibility for efficient Virtual Machine (VM) migration. The algorithm dy-namically predicts overloaded VMs, minimizes migration time, and adapts to traffic fluctuations. Experimental results demonstrate a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{30 \%}$</tex> reduction in migration time and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{25 \%}$</tex> improvement in resource utilization compared to Multipath TCP (MPTCP) and heuristic methods. Key innovations include: •A Decentralized Partially Observable Markov Decision Process (Dec-POMDP) framework for scalable multi-agent coordination. •Integration of actor-critic algorithms with epsilon-greedy exploration and parameter sharing for robust policy learning. •SDN-driven global state awareness to optimize VM placement.