Towards optimal edge resource utilization: Predictive analytics and reinforcement learning for task offloading

Edge computing brings computation closer to the user devices. This proximity allows user devices with limited resources to execute complex and computation-intensive tasks on the edge, thus minimizing the latency in task execution. A typical edge device employs its resources for executing user tasks. When the resources are unavailable, the tasks are offloaded either to the network’s other edge devices or to the cloud. Selecting such an optimal location and adhering to various task requirements is a multiple-constraint optimization problem, and several works exist to addressing this issue. However, the majority of these works emphasize on determining the optimal location to offload tasks based on current resource availability and edge intelligence. In this work, we aim to predict the resource availability at the edge device before the task offloading decision is made. Further, we design a leader-based optimal task offloading decision approach where the edge-server uses a reinforcement learning model to determine such optimal location. We formulate a dynamic optimization problem to solve the optimal location issue in edge computing environment. The proposed reinforcement learning based technique maximizes the system utility by optimizing task offloading and resource allocation policies. Through experimental results, we show that our model achieves efficient edge resource utilization and better individual profit for edge devices. The results also show that resource idle time on edge devices is significantly lower as compared to the non-predictive models.