Multi-Objective Reinforcement Learning For Service Selection In Multi-Cloud Iot Environments

Multi-cloud computing, where multiple cloud service providers are utilized, enhances resilience and reduces dependency on a single provider. However, selecting appropriate cloud services for IoT applications with strict Quality of Service requirements remains a complex, NP-hard problem. This paper proposes Multi-Objective Deep Reinforcement Learning (MODRL) approaches to optimize the multi-cloud service selection process. MODRL can handle cloud environments' dynamic and high-dimensional nature by adapting to changing resource availability and performance metrics. We adapted, implemented, and tested two DRL methods: Pareto Conditioned Network (PCN) and Pareto Q-Learning (PQL). The results are promising, demonstrating the effectiveness of these methods in optimizing service selection in real-time, dynamic cloud environments, especially the MODRL approach, PCN, which showed the best results.