Constrained Multi-Objective Deep Reinforcement Learning for Resource Allocation in Edge–Cloud IoT Networks

Abstract

The rapid proliferation of Internet of Things (IoT) applications has created heterogeneous, bursty and deadline-sensitive workloads that cannot be managed efficiently through static edge–cloud allocation. This study proposes a constrained multi-objective deep reinforcement learning (CMO-DRL) architecture for jointly selecting task placement and computational capacity across IoT devices, edge servers and cloud infrastructure. The design combines an actor–critic scheduler, normalized latency–energy–cost objectives, deadline constraints, action masking and a Pareto-aware decision layer. A mathematical formulation, implementation workflow, evaluation protocol and reproducible illustrative scenario are presented. The numerical examples demonstrate how a candidate policy could be evaluated against first-come-first-served, edge-first and latency-only reinforcement learning baselines, but are not claimed as outcomes of an executed experiment. The framework emphasizes feasibility, interpretable trade-offs, scalable orchestration and reproducible validation under variable traffic intensity.