| Mobile edge computing is a computing paradigm that is coming to life.Emerging 5G communication technologies and networking techniques organically organize edge networks such as the inherently decentralized,self-organizing and autonomous Internet of Things(Io T),heterogeneous networks(Het Nets)and UAV(Unmanned Aerial Vehicle)networks to divide and collaborate on large-scale,complex tasks.Edge terminals need local autonomous decision making,but due to their limited computing power and storage resources,they need to offload task and resourceaware information to edge networks or idle mobile devices for collaborative computing,and return the computation results to the source devices to complete the task computation.However,such dynamic time-varying edge collaborative computing faces unprecedented challenges:(1)the challenge of reliable dynamic communication spectrum resource sensing and effective utilization?(2)the challenge of optimizing trusted real-time collaborative computing in dynamic time-varying mobile networks?(3)the challenge of susceptibility to malicious interference during task offloading transmission? and(4)the challenge of reliable and efficient joint optimization of task computation offloading and resource allocation.This thesis will address the above four challenges in depth.First,this thesis investigates the problem of dynamic communication spectrum resource sensing and effective utilization during task transmission in mobile edge networks.In mobile edge network communication and collaborative computing,the presence of interference can cause false alarms or fail to sense the channel access status of primary user in time,leading to improper decisions about spectrum access by secondary users resulting in poor spectrum resource utilization,thus affecting the communication and collaborative computing performance of the entire edge network.First,a primary-priority dynamic spectrum resource allocation model is proposed to model the resource interaction process between primary and secondary users as a Markov decision process(MDP),and the optimal spectrum allocation strategy for secondary users is derived,and the MDP is used to capture the evolution of the timing characteristics of dynamic spectrum,so as to achieve the priority queuing of primary and secondary users.The optimal tradeoff benefit of spectrum fairness and throughput can be obtained according to the spectrum access strategies of sub-users under different optimal criteria.Secondly,a dynamic spectrum access algorithm based on recursive deep reinforcement learning is proposed,which allows secondary users to modify their parameters to select the optimal spectrum access policy and uses a Dueling Deep Q-Network(Dueling DQN)with prioritized empirical replay combined with a recursive neural network to improve the convergence speed and prediction accuracy of the algorithm.Second,this thesis investigates the optimization problem of real-time offloading decisions in dynamic time-varying mobile edge computing networks.To address the challenge that computational offloading in the wireless fading environment in MEC is computationally complex in large-scale network tasks and easily blocked by malicious seizure of edge resources,this thesis proposes a Policy-based trusted collaborative computing model.To reduce the computational complexity of the algorithm,a deep reinforcement learning algorithm with joint offloading action quantization and attention mechanism is proposed to approximate the continuous action values to a finite number of discrete values.Due to the task latency constraint,there exist some mobile devices that are unable to complete the collaborative computation offloading within the specified time,which means that generating high-dimensional offloading decision actions in each time frame is not only inefficient but also unnecessary,so an order-preserving pruning strategy is used to prune the infeasible offloading decisions to reduce the computational complexity of the algorithm and thus achieve efficient computational performance while ensuring accuracy.Third,this thesis investigates the malicious interference problem in cooperative computing in mobile edge networks and proposes a Stackelberg game anti-interference cooperative computing model,which accurately describes the interference game relationship among source devices,cooperative computing devices and interferers in cooperative computing in edge networks.In this thesis,the attack-defense confrontation between multiple mobile devices and jammers is modeled as a Stackelberg game,where the source device is the leader,the co-computing device acts as a subleader and adjusts its anti-interference strategy according to the strategy of source device to improve the communication anti-interference performance of source device,and the jammers act as followers.A Stackelberg Game Policy Learning(SGPL)algorithm for deep reinforcement learning is subsequently designed,where the leaders(co-computing devices)update their training parameters using the total derivatives of the objective function,while the followers(i.e.,jammers)update their training parameters using an independent gradient dynamics policy.Loops are eased and convergence is accelerated by introducing differential dynamics in the leader training network to reflect the interaction structure of the critic and actor network layers.Finally,this thesis investigates the problem of joint optimization of collaborative computing and resource allocation in mobile edge networks.To address the coupling problem in the joint optimization of collaborative computing offloading decision and resource allocation,this thesis proposes a Differential Dynamic Gradient Descent(DDGD)optimization algorithm,which decomposes the coupled joint optimization problem with inequality constraints and equation constraints into two network layers and constraint functions are integrated into a larger end-to-end training network.These two-layer networks encode dependencies and optimization constraints between parameter hidden states that cannot be captured by numerical optimization models or fully connected layer neural networks.Thanks to the circular and self-repeating learning structure and the information stored in the differential dynamics network,the proposed algorithm enables better and more intelligent decisions in searching for solution trajectories without pre-setting the exact parameters and reduces the complexity of the network. |
