Confidence-Gated Adaptive Routing (CGAR): A Bandit-Theoretic Approach to the Cold-Start Problem in Reinforcement-Learning-Based API Gateways
Original Research
Volume 14, Issue 4, (2026) Cite this article
Submit your manuscript
Original Research
Volume 14, Issue 4, (2026) Cite this article
Submit your manuscript
Khushi Mittal 
Pages: 37-43
Recent work applies deep reinforcement learning (RL) to API gateway traffic routing in cloud microservice architectures, reporting consistent gains over static heuristics once the underlying policy converges. A persistent gap in this literature is the omission of any evaluation of the transient learning period that follows deployment, autoscaling, or backend pool changes, referred to here as the cold-start problem. This paper introduces Confidence-Gated Adaptive Routing (CGAR), a framework that maintains a per-backend confidence estimate and uses it to continuously blend a static heuristic with a learned policy on a per-backend basis, re-triggering safe routing individually for newly added backends without resetting global learning. The problem is formalised as a non-stationary, per-arm warm-start bandit problem, and CGAR is evaluated against seven baselines over 20 independent seeds using a discrete-event simulator subjected to four scheduled backend pool churn events. CGAR attains a total regret of 0.8×103, reducing regret by 81.6% relative to Pure DQN (p < 0.001), 80.4% relative to Eps-Greedy DQN (p < 0.001), and 91.9% relative to Round Robin (p < 0.001), while sustaining a mean SLA violation rate of 0.3%. A churn-window versus steady-state regret decomposition further shows that CGAR limits churn-window regret to 0.3×103, substantially lower than every reinforcement-learning baseline, demonstrating that the gating mechanism provides targeted protection precisely when cold-start costs are highest. Ablation results are reported transparently: a fixed high-weight heuristic blend attains marginally lower regret than the adaptive variant in the present simulator, and this trade-off is discussed rather than concealed. CGAR nonetheless retains practical value over static configurations because it re-triggers cold-start protection automatically per backend without manual threshold tuning, and it outperforms every learning-based baseline with strong statistical significance. The work contributes a formal connection between warm-start bandit theory and API gateway routing that is absent from the current literature.
Reinforcement Learning, Multi-Armed Bandits, Api Gateway, Microservices, Cold-Start Problem, Traffic Routing, Cloud Computing