Research On Automated Structure Optimization Technology Of Edge Intelligence Models Based On Deep Reinforcement Learning

Cloud computing has the ability to process large-scale complex data,which has led to the widespread application of deep learning technology in various industries and has achieved significant achievements.However,with the booming development of the Internet of Things(Io Ts),traditional cloud computing cannot meet the real-time and security requirements of diversified and massive Io Ts data near the edge side only through centralized cloud servers.Therefore,the edge computing paradigm has been proposed to provide real-time,secure,and reliable computing services.In order to deploy Deep Neural Network(DNN)models to edge devices to provide intelligent services,Edge Intelligence has rapidly developed.Among them,optimizing the structure of DNN models to alleviate the storage and computing resources required for edge intelligence services is a very promising solution,leading to the emergence of various DNN model structure optimization techniques.Based on whether expert manual optimization of the model is required,optimization techniques can be classified into two categories: manual model structure optimization and automated model structure optimization.The manual model structure optimization method not only requires domain expert knowledge but also incurs a large amount of expert manpower and time costs.Therefore,this article mainly focuses on automated model structure optimization technology.Automated model structure optimization techniques mainly include Automated Model Compress(AMC)and Neural Architecture Search(NAS).Existing AMC and NAS technologies are usually used to compress or search for a model that meets the constraints in simple scenarios with single edge devices and single datasets.However,for complex multi-task model deployment scenarios,it is necessary to perform repeated model structure optimization separately for each task scenario,which is an extremely time-consuming and laborious process.Therefore,how to achieve automated model compression and neural architecture search in multi-task scenarios has become a key issue in the field of edge intelligence model deployment.To address the above-mentioned issues and challenges,this paper proposes a multitask model optimization framework based on Importance Weighted Actor-Learner Architecture(IMPALA)for both AMC and NAS.The framework first establishes a Markov Decision Process(MDP)model in a Reinforcement Learning(RL)setting for specific structural optimization problems,and then designs a correct and efficient policy optimization algorithm to train the agent.The detailed research work and innovative points are as follows:1.To address the multi-task automated model compression problem,this paper proposes the Multi-Task Automated Channel Pruning framework.The paper first analyzes in detail the principles of automated channel pruning and the key points of channel pruning in a multi-task scenario.It then defines the multi-task automated channel pruning problem as a constrained optimization problem and establishes an MDP model,including the state of the channel pruning environment,the actions of the agent,and the reward function for guiding policy optimization.A policy optimization algorithm based on V-trace value estimation is designed.Experimental results show that MTACP has the ability to solve the multi-task automated channel pruning problem and has significantly improved pruning speed compared to the latest existing automated pruning methods.2.To address the multi-task neural architecture search problem,this paper proposes the Multi-Task Neural Architecture Search framework.The paper first analyzes in detail the principles of multi-objective neural architecture search and decides to use a performance evaluation method based on weight sharing to achieve efficient multi-objective architecture search.The multi-task multi-objective neural architecture search problem is then defined as a constrained optimization problem and an MDP model is established,including the state of the architecture search environment,the actions of the agent in selecting sub-architectures,and the reward function containing multiple optimization objectives.A policy optimization algorithm based on Monte Carlo sampling value estimation and Importance Sampling techniques is designed.Experimental results show that MTNAS has the ability to solve the multi-task neural architecture search problem and has similar performance compared to expert-designed models and models searched by other latest NAS methods.However,the search efficiency of MTAS has significantly improved compared to the latest NAS methods.