Task offloading and resource allocation algorithm based on deep reinforcement learning for distributed AI execution tasks in IoT edge computing environments

Recently, the application of Artificial Intelligence (AI) in the Internet of Things (IoT) devices is increasing. As these devices are limited in processing and storing massive computations of AI applications, researchers are searching for methods to overcome these limitations. One of these applications is Convolutional Neural Network (CNN) processing, which is common in object detection and image classification. A neural network consists of layers with a large number of neurons and requires high processing power to run. A CNN can be partitioned into segments and offloaded as tasks of IoT devices to cloudlet servers on the edge. By utilizing the edge servers available in the environment, the total latency in the system and energy consumption by IoT devices can be optimized. Making decisions about offloading CNN segmented layers and allocating resources to each of them is the challenge. In this paper, we propose a method based on deep reinforcement learning that divides the offloading and resource allocation problem into two minor problems. This algorithm updates the offloading policy based on information from the environment, and with the help of the Salp Swarm Algorithm (SSA), optimizes resource allocation. The proposed method is tested for different deep-learning tasks of IoT devices under different capacities of cloudlet servers. The simulation results show the proposed algorithm has the least cost in terms of latency and power consumption and on average has improved 92%, 17%, and 12% compared to full local, full offload, and Jointly Resource allocation and computation Offloading PSO (JROPSO) methods respectively.