Reducing Digital $\text{CO}_{2}$ Footprint in IT Systems Using Temporal Difference Learning for Energy-Aware Scheduling

Optimizing resource allocation is critical for reducing power consumption and minimizing carbon emissions in high-performance computing environments. This research presents a Deep-Q-Network based Reinforcement Learning DQNAgent to enhance decision-making in CPU/GPU allocation, and workload allocation. For better decision making, the proposed DQNAgent follows an epsilon-greedy policy to balance exploration and exploitation. The agent is trained for 1000 episodes, keeping exploration rate (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\epsilon$</tex>) decreases from 1.0 (random action selection) to 0.01 (optimal policy execution), which enables the DQNAgent to make more informed decisions. The DQNAgent is trained on a CPU/GPU workstation. The experimental results clearly indicate that training the agent on a CPU-based system to learn an optimal policy to minimize emissions takes approximately 7 hours and 17 minutes, while utilizing a GPU workstation significantly reduces execution time to less than 2 minutes, thereby accelerating real-time decision-making. Results indicate that the proposed method effectively optimizes energy efficiency, reducing overall power consumption while maintaining computational performance. This research work paves way for leveraging Artificial Intelligence driven reinforcement learning for sustainable and intelligent computing systems.