<p>The escalating adoption of cloud computing among internet users and the rapid expansion of cloud data centers have brought about new challenges for cloud service providers. These challenges encompass issues such as heightened energy consumption in data centers, increased energy expenses, and carbon taxation. One viable solution to address these challenges is the optimal placement of virtual machines. It is important to note that virtual machine placement falls under the category of NP-Hard problems, making it particularly complex. Consequently, researchers have explored various approaches, including heuristics, meta-heuristics, and more recently, reinforcement learning-based strategies, to tackle this issue. Given the dynamic nature of the problem’s state space, reinforcement learning has shown superior performance compared to heuristics and meta-heuristics. In the domain of virtual machine placement, numerous reinforcement learning-based approaches have been introduced, each with distinct objectives such as reducing energy consumption, minimizing virtual machine migrations, and enhancing data center productivity. However, these methods often grapple with challenges related to the complexity of the state space and relatively slow convergence rates. To address these limitations, there has been a growing interest in leveraging deep learning and deep reinforcement learning techniques. These approaches offer the promise of expedited learning. In this research, we present an approach that combines reinforcement learning and deep reinforcement learning to mitigate energy costs and carbon emissions. By harnessing the adaptability of reinforcement learning within dynamic environments and accelerating learning through deep reinforcement techniques, we aim to achieve a reduction in both energy consumption and virtual machine migration rates. The results of the article, which is based on Double-Deep Q-learning, were compared with different methods for creating 14,000 virtual machines between nine data centers in different geographical areas, each of which contains 90 hosts. Regarding the energy consumption index, our proposed method was able to reduce consumption by 16%, 41%, and 24% compared to Q-learning, heuristic, and meta-heuristic approaches, respectively. Regarding the indicators of carbon cost, carbon tax and total cost, which includes energy and carbon cost together, our proposed method has obtained better results in all cases. One of the important indicators in the correct placement of virtual machines is the reduction of their migration over time. In this index, the proposed approach has been able to reduce migration by 15% and 27% compared to Q-learning and meta-heuristic approaches.</p>