Optimization of fog resource allocation in IoT using Deep Reinforcement Learning

Fog computing is an emerging paradigm that extends the cloud concept to the edge. It provides computing, storage, control, and networking capabilities for realizing the Internet of Things (IoT) applications. In the fog computing concept, the IoT devices offload its data or computationally expensive tasks to the fog nodes within its proximity, instead of distant cloud. In this paper, we address the problem of optimal allocating the limited resources of fog nodes to the IoT applications. Current approaches of fog resource allocation are not sufficiently adaptable in noisy and uncertain environments. Using learning-based algorithms is therefore essential. Resource allocation problem can be considered as an online decision-making system where fog nodes should decide whether processing locally the receiving requests from IoT devices or sending them to distant cloud nodes. We model the fog resource allocation problem as a Markov decision process and solve it by the Deep reinforcement learning approach. Based on policy-gradient algorithm, fog nodes learn how to schedule the IoT tasks in an optimal way . The proposed method is compared with the non-learning approach in which tasks are assigned to fog nodes based on their length and without the consideration of task priority. The obtained results, according to the cumulative reward during the implementation process of the proposed algorithm, indicate that the resource allocation policy has been learned online. This improves the average slowdown and average slowdown in difficult conditions for a system with different task priorities, when compared to the non-learning method.