Policy separated learning assisted COPRAS based leader election for dynamic clustered distributed systems

Leader election is a fundamental coordination mechanism in clustered distributed systems operating under workload variability, node churn, and probabilistic failures. Existing approaches are typically either deterministic, offering limited adaptability under dynamic conditions, or fully adaptive schemes that increase coordination overhead and reduce decision stability. Static multi-criteria decision-making approaches further rely on fixed re-election policies that cannot effectively balance leader quality and coordination efficiency in non-stationary environments. This paper introduces policy-separated leader election, a framework that decouples deterministic candidate ranking from adaptive re-election control. Based on this principle, we propose LACOPRAS, a learning-assisted leader election framework that integrates COPRAS-based multi-criteria ranking with lightweight adaptive policy regulation. COPRAS is preserved as an interpretable deterministic ranking mechanism, while learning is restricted to adaptive criterion weighting and policy-level regulation of re-election timing. The framework is evaluated using trace-driven discrete-event simulations under dynamic operating conditions and compared with deterministic, heuristic, adaptive MCDM, and reinforcement learning baselines. Experimental results show that LACOPRAS achieves an average leader utility of 0.7840 while reducing average coordination overhead to 39.22 compared with 46.00 for Adaptive TOPSIS and 49.00 for RL-only coordination. The results further demonstrate stable adaptive coordination trends and scalable overhead behavior across the evaluated scenarios.