مسعود ربيعي

With the rapid growth of deman‎d for cloud computing services, the number of data centers has increased worldwide, an‎d consequently, energy consumption an‎d greenhouse gas emissions have intensified. Therefore, energy consumption optimization has become one of the central challenges of cloud computing. Energy-aware workflow scheduling is a multi-objective optimization problem that simultaneously considers energy consumption an‎d the length of the scheduling period (Makespan); an inherently NP-Hard problem that, with the dynamics an‎d high scale of cloud environments, limits the efficiency of traditional an‎d meta-heuristic methods. In this research, a new approach for energy-aware workflow scheduling based on deep reinforcement learning is presented. The core of the proposed method is the Noisy Double Dueling Deep Q-Network architecture, which reduces the overestimation bias in valuation by using Double Q-Learning, an‎d replaces classical strategies such as ε-Greedy with state-dependent intelligent exploration with noisy layers. The reward function also optimizes three key components—energy consumption, scheduling period length, an‎d deadline compliance—integrally. The efficiency of the method has been eva‎luated on a simulation platform an‎d on real scientific workflows with different scales. The results show that the proposed approach outperforms other deep reinforcement learning methods in reducing energy consumption as well as in algorithm behavioral indicators—including convergence speed an‎d stability. The introduced hybrid architecture provides an efficient an‎d practical solution to the workflow scheduling problem in dynamic cloud environments an‎d can be a basis for the development of green an‎d scalable schedulers in future generations of cloud computing.

رايانش ابري , زمان‌بندي گردش‌كار , بهينه‌سازي مصرف انرژي , زمان‌بندي آگاه از انرژي , يادگيري تقويتي عميق , شبكه‌هاي عصبي عميق , الگوريتم DQN

Cloud computing , Workflow Scheduling , Energy Optimization , Energy-Aware Scheduling , Deep Reinforcement Learning , Deep Neural Networks , DQN Algorithm

Masoud Rabiei

Dr. Marzieh Malekimajd