Deep Reinforcement Learning Empowered Destination Driven Computation Offload Mechanism

As the amount of information and the computational complexity of applications in the Industrial Internet of Things increase,resource-constrained devices increasingly rely on computational offloading technology.Computation offload technology can transfer the computing requirements and data generated by lightweight devices in the Internet of Things to computing offload nodes with sufficient computing resources.On the one hand,the lightweight device can save valuable computing,storage and energy resources,and on the other hand,it can also achieve higher delay requirements by calculating the high computing force of nodes.The computational unload decision algorithm is the core of the computational unload technology,but in the emerging destination-driven computational unload scenario,the traditional computational unload algorithm cannot make effective unload strategy.Therefore,it is of great significance for the development of IoT application under this model to study the computing offloading decision mechanism under the destination-driven computing offloading scenario.In this paper,the corresponding computing and unloading decision mechanism is designed for two destination-driven computing unloading IoT scenarios with special requirements:(1)In view of the limited energy,IoT scenarios on the dynamic change of network environment,this paper puts forward distributed goal-driven calculation based on the depth of intensive study uninstall decision-making mechanism(D~3COM),will be combined with the depth of the deep learning and reinforcement learning reinforcement learning is introduced into the target driver unload decision model,through the neural network generalization function and memory and solve the problem that the computing node locality,through the operation mode of the centralized training distributed execution(CTDE)lighten the burden of training and storage nodes,improve the network convergence speed.Finally,through the data recorded during the interaction between the agent and the environment,the relationship between the calculation unloading strategy and the multi-dimensional characteristics such as network state and task state can be learned.Experimental results show that D~3COM achieves good results in reducing task delay,balancing energy consumption distribution and prolonging the life cycle of computing network.(2)According to different delay sensitive degree of goal-driven uninstall scenario,based on the theory of mixed expert model(Mo E),Mo E module,introduced D~3COM characteristics in the process,establish more child network study different mapping relations hidden in the data,and the Manager network adaptive learning between sensitive features and how to combine multiple child network connection.Through the Mo E module,the simple one-hot feature is transformed into a low-dimensional dense feature with high characterization ability.Experimental results showed that Mo E-based D~3COM model achieved excellent performance in formulating differentiated computing offloading strategies based on task sensitivity.