Reliable Computation Offloading of DAG Applications in Internet of Vehicles Based on Deep Reinforcement Learning

To solve the partial offloading problem of Directed Acyclic Graph (DAG) applications in the Internet of Vehicles (IoV) under edge computing environment, this paper proposes a partial offloading algorithm for DAG application via deep reinforcement learning. The proposed algorithm considers the complexity of server computing resource allocation and computing offloading decision to maximize the comprehensive utility of time delay and energy consumption. First, we propose a partial offloading model of DAG application. In order to determine the execution priority of each sub-module in DAG application, an execution priority algorithm is designed to transform DAG application into sequence structure. Then, a recurrent neural network-based sequence-to-sequence network is designed as a strategy network, and a partial offloading algorithm based on deep reinforcement learning is proposed. Finally, to ensure the reliable offloading of DAG application in the IoV, a re-decision scheme is proposed in the case of edge server failure. Experimental results show that the proposed algorithm achieves better comprehensive utility than the baseline algorithms under the same computation offloading scenario, and the designed fault handling scheme can ensure reliable offloading of DAG applications in the case of server faults.