ADAPTIVE CDN NODE SELECTION IN DYNAMIC ICT SYSTEMS USING ONLINE CONTROLLED EXPERIMENTS AND CHANGE DETECTION MULTI ARMED BANDIT ALGORITHMS

Modern info‑communication technology (ICT) infrastructures such as content‑delivery networks (CDNs) must continuously tune low‑level parameters to deliver high performance under variable and non‑stationary network conditions. This paper investigates how online controlled experiments—including classical A/B tests and adaptive multi‑armed bandit (MAB) algorithms—can be used to optimise CDN node selection. We formalise the optimisation problem as minimising a network‑performance objective of average latency, one of key metrics used to measure network performance. After reviewing prior work on A/B testing and MABs, we propose a new Change‑Detected Upper‑Confidence‑Bound (CD‑UCB) algorithm that couples the classical UCB arm‑selection rule with a cumulative‑sum (CUSUM) change‑detection statistic. The CD‑UCB algorithm rapidly resets its estimates when performance shifts, enabling faster adaptation to non‑stationary environments. A simulation of CDN node selection with three nodes having different latency distributions is used to compare four approaches: simple A/B testing, sequential A/B testing with early stopping, standard UCB, and the proposed CD‑UCB. Each algorithm is evaluated using cumulative regret , rolling average latency, rolling throughput and the percentage of requests sent to the optimal node. In a stationary setting, all methods eventually identify the best node, but MAB‑based approaches converge faster and exhibit lower regret. When the environment changes abruptly, simple and sequential A/B tests fail to adapt and incur high regret, whereas standard UCB adapts slowly. CD‑UCB detects changes quickly and nearly matches the instantaneous optimal policy, achieving the lowest cumulative regret and closely tracking the true optimal latency. The results demonstrate that adaptive MAB algorithms with change detection are better suited than static A/B tests for optimising dynamic ICT infrastructures. The study concludes with recommendations for applying MAB‑based online experiments to infrastructure optimisation and suggests future work on multi‑objective optimisation, contextual bandits and evaluations on real‑world testbeds.