SAGE: A Self-Adaptive Guided Explainable Scheduler for Heterogeneous Computing Environments Using Deep Reinforcement Learning and Digital Twin Predictors

Abstract

Task scheduling in heterogeneous computing environments remains a critical challenge, requiring simultaneous optimisation of multiple competing objectives including execution latency, energy consumption, monetary cost, and Service Level Agreement (SLA) compliance. Traditional heuristic schedulers such as Round-Robin and Min-Min lack adaptability to dynamic workload conditions, while black-box reinforcement learning (RL) agents lack the transparency required for trustworthy deployment in production systems. This paper presents SAGE (Self-Adaptive Guided Explainable Scheduler), a novel framework that synergistically combines three components: (1) a Proximal Policy Optimisation (PPO) policy trained via deep reinforcement learning for candidate action proposal, (2) a Digital Twin (DT) predictor ensemble based on Gradient Boosting Regressors for multi-objective evaluation and re-ranking of candidates, and (3) a SHAP-based explainability module that generates human-readable, contrastive justifications for every scheduling decision. SAGE further incorporates a self-adaptation mechanism that continuously retrains the Digital Twin on observed outcomes, enabling the system to handle concept drift caused by workload variations and resource degradation. Evaluation against five baseline schedulers across 30 experimental episodes demonstrates that SAGE achieves the lowest SLA miss rate of 28.00% (compared to 48.17% for Random and 44.50% for Round-Robin), the highest cumulative reward of −31.21 (a 40.4% improvement over the next-best baseline), and competitive latency of 5.96 seconds. These results confirm that integrating RL-based proposal, DT-based look-ahead evaluation, and explainable decision-making produces a scheduler that is simultaneously adaptive, efficient, and interpretable.