🤖 AI Summary
To address low payment routing success rates and insufficient system resilience under dynamic gateway performance, this paper proposes a control-theoretic adaptive dynamic routing framework. Methodologically, it integrates generalized feedback control, reinforcement learning, and multi-armed bandit techniques to construct a closed-loop routing controller: departing from conventional PID designs, it introduces a novel feedback mechanism that enables gateway scores to adaptively converge toward their true success rates, balancing short-term responsiveness with long-term stability. The key contribution lies in deeply embedding control theory into payment routing decisions, enabling real-time, performance-aware closed-loop operation—comprising performance sensing, feedback-driven adaptation, and policy optimization. Empirical evaluation in an online production environment demonstrates that the proposed framework improves transaction success rate by up to 1.15% over rule-based routing, significantly enhancing system reliability and elasticity.
📝 Abstract
This paper introduces a control-theoretic framework for dynamic payment routing, implemented within JUSPAY's Payment Orchestrator to maximize transaction success rate. The routing system is modeled as a closed-loop feedback controller continuously sensing gateway performance, computing corrective actions, and dynamically routes transactions across gateway to ensure operational resilience. The system leverages concepts from control theory, reinforcement learning, and multi-armed bandit optimization to achieve both short-term responsiveness and long-term stability. Rather than relying on explicit PID regulation, the framework applies generalized feedback-based adaptation, ensuring that corrective actions remain proportional to observed performance deviations and the computed gateway score gradually converges toward the success rate. This hybrid approach unifies control theory and adaptive decision systems, enabling self-regulating transaction routing that dampens instability, and improves reliability. Live production results show an improvement of up to 1.15% in success rate over traditional rule-based routing, demonstrating the effectiveness of feedback-based control in payment systems.