Autonomous computation offloading and auto-scaling the in the mobile fog computing: a deep reinforcement learning-based approach

The Fog Computing (FC) paradigm is rapidly becoming an appropriate framework for the infrastructure related to the Internet of Things (IoT). FC can be a good framework for mobile applications in the IoT. This architecture is referred to as the Mobile Fog Computing (MFC). Modules in the applications can be sent to the Fog or Cloud layer in the event of the lack of resources or increased runtime on the mobile. This increases the efficiency of the whole system. As data is entered sequentially, and the input is given to the modules, the number of executable modules increases. So, this research was conducted to find the best place in order to run the modules that can be on the mobile, Fog, or Cloud. According to the proposed method, first, the Fog Devices (FDs) were locally evaluated using a greedy technique; namely, the sibling nodes followed by the parent and in the second step, a Deep Reinforcement Learning (DRL) algorithm found the best destination to execute the module so as to create a compromise between the power consumption and execution time of the modules. The evaluation results obtained regarding the parameters of the power consumption, execution cost, delay, and network resource usage showed that the proposed method on average is better than the local execution, First-Fit (FF), and standard DRL by 18, 6, and 2%, respectively.