CATS-DRL-PFOA: A Cost-Aware Task Scheduling Framework using Deep Reinforcement Learning and Pufferfish Optimization in Secure Cloud Environments

Cloud computing has become the dominant paradigm for elastic, on-demand computing services, but efficient and secure task scheduling across heterogeneous virtual machines (VMs) and data centers remains a critical challenge for cloud service providers (CSPs). Existing schemes often optimize only a subset of objectives, such as response time or energy, and ignore SLA-violations, trust, and dynamic workload patterns. Motivated by the workload, cost, and SLA models described in the reference chapter on host and VM workload estimation, this paper proposes CATS-DRL-PFOA (Cost-Aware Task Scheduling using Deep Reinforcement Learning and Pufferfish Optimization Algorithm), a hybrid framework that integrates deep learning, reinforcement learning, and swarm-based metaheuristics under a cloud security and Generative-AI-assisted SLA environment. The proposed method uses LSTM-based workload prediction and CNN-based anomaly detection for VM health, Q-learning-based task scheduling policies, and a Pufferfish Optimization Algorithm (PFOA) for selecting cost-optimal, SLA-compliant VM–task mappings. Service Level Agreement (SLA) metrics such as machine availability, success rate, and response efficiency are modeled using and extending equations (7.8)–(7.11) from the base chapter. We generate a synthetic dataset reflecting realistic VM capacities, task sizes, and electricity costs and conduct extensive simulations under 100–500 task loads. Results show that CATS-DRL-PFOA reduces average response time by up to 2.43×, improves SLA compliance by 7.1%, and decreases CSP cost by 31.2% compared with a heuristic SMA-based baseline. In addition, the LSTM predictor achieves 94.3% accuracy in workload forecasting and the CNN-based anomaly detector reaches an F1-score of 0.940. These improvements demonstrate that tightly coupling deep reinforcement learning with cost-aware metaheuristic optimization yields robust, secure, and efficient cloud task scheduling suitable for modern AI-driven data centers. 

Keywords: Cloud computing, task scheduling, deep reinforcement learning, LSTM, Pufferfish Optimization Algorithm, SLA, cloud security, Generative AI, NLP.