<![CDATA[Round-trip time (RTT) is a key latency indicator for quality-of-service (QoS) control and task orchestration in cloud–edge systems. However, RTT is highly time-varying due to congestion dynamics, routing changes, and fluctuating traffic conditions, motivating short-term prediction to enable proactive decision making. This paper investigates a hybrid neuro-fuzzy baseline for RTT prediction implemented using an Adaptive Neuro-Fuzzy Inference System (ANFIS) with subtractive-clustering-based initialization to avoid rule explosion in high-dimensional inputs. A controlled dataset was generated in Mininet using a dumbbell topology with injected delays (1–1000 ms). In total, 100,000 raw RTT records were collected (100 RTT measurements per run across 1000 runs) and aggregated into 1000 supervised samples paired with TCP-state features. Experiments followed a unified and reproducible protocol with a fixed 60/10/30 train/validation/test split, train-only feature standardization, train-only target normalization with inverse transformation for reporting, and validation-based checkpoint selection. The ANFIS baseline (radius ????=0.5r=0.5, 19 rules) achieved RMSE/ MAE/ MAPE/ ????2 of 1191.63/ 751.02/ 0.001921/ 0.999996 on validation and 1207.23/ 664.70/ 0.001311/ 0.999996 on testing. Training required 546.91 s, while inference remained lightweight (0.0846 s for 100 validation samples and 0.1493 s for 300 test samples). Diagnostic analyses using learning curves, parity plots, residual inspection, and empirical error distributions further supported the strong agreement between predicted and observed RTT values. These results indicate that ANFIS with subtractive clustering can deliver accurate and low-latency RTT prediction suitable for QoS-aware orchestration pipelines where training can be performed offline.]]>
Copyrights © 2026
