Guarded deep reinforcement learning for SLA latency and energy aware task scheduling in IoT edge cloud environments

With the fast-growing IoT applications, there is a need for intelligent task-scheduling mechanisms that can meet the latency, energy, and service-level agreement (SLA) constraints in a dynamic edge–cloud environment. This may not be met with traditional heuristic and deep reinforcement learning (DRL)-based schedulers under resource heterogeneity, workload volatility, and varying network conditions. In this paper, we propose a guarded deep reinforcement learning framework, called TriSchedRL, for SLA-aware, latency-aware, and energy-aware IoT task scheduling. The framework consists of three key components: (i) latency, energy and SLA-risk prediction modules are added to improve system context awareness; (ii) a guard layer filters scheduling actions and validates them before execution to avoid infeasible resource allocations; and (iii) an adaptive tri-objective reward mechanism dynamically balances SLA compliance, latency and energy consumption. TriSchedRL is evaluated through a discrete-event simulation on edge-cloud systems with varying workload intensities, bursty traffic patterns, heterogeneous resource configurations, and network conditions. Experimental results show that TriSchedRL consistently outperforms heuristic, fixed-weight DRL, and recent constrained reinforcement-learning baselines. TriSchedRL has an SLA violation rate of 3.0% for the 10,000-task workload with an average latency of 45·6 ms and energy consumption of 6·0J/task. The proposed framework is further demonstrated to be effective and robust through statistical significance analysis, component-level ablation studies, and adaptive-weighting experiments. The results demonstrate the feasibility of using TriSchedRL for intelligent task scheduling in an IoT-enabled edge–cloud system. The code implementation is available at https://github.com/AngularWaseem/TriSchedRL .