Abstract
Software-Defined Networking (SDN) enables centralised and programmable traffic control, but adaptive optimization in operational networks remains challenging when safe online exploration is limited and only historical controller traces are available. This study proposes a log-driven Proximal Policy Optimisation (PPO) framework for adaptive SDN traffic control that learns directly from recorded state–action–reward transitions. The method uses a replay-based pseudo-environment constructed from controller logs. It combines clipped PPO updates with action-consistency regularisation and running state normalisation to improve stability under logged-data constraints. The empirical evaluation shows that the learned model reconstructs the dominant response pattern observed in the traces, preserves a positive relationship between the principal control-related predictor and the response, and reveals a non-uniform interaction structure across telemetry features. The framework also differentiates systematically across operating conditions and experimental groups, with category means ranging from 0.78 to 1.24 and group medians ranging from −0.12 to 1.12, while receiver operating characteristic analysis yields an area under the curve of 0.714. The practical network evaluation further shows that the PPO-controlled setting improves overall throughput, packet loss, jitter, and flow-completion success relative to the baseline controller. These results indicate that log-driven, stability-constrained PPO can provide a stable and informative basis for adaptive SDN traffic control when policy learning must rely on historical controller data rather than unrestricted live-network experimentation.