Research On Computation Offloading For Reusable Tasks In Internet Of Things

The computation offloading based on Multi-access Edge Computing(MEC)has been widely taken as one of the supporting technologies for various newly emerging Internet of Things(IoT)applications.The new IoT application can be composed of a set of modular,reusable tasks to achieve flexibility.By fully utilizing the characteristics of task data reuse,it can further reduce the amount of offloaded data and improve the performance of computation offloading.Therefore,it has great theoretical value and application prospects to deeply research the computation offloading for reusable tasks in IoT.From three perspectives including vertical resource collaboration,horizontal device cooperation,and long-term dynamic optimization,this thesis conducts research on collaborative computation offloading based on cache-assisted MEC,cooperative computation offloading based on coalitional game theory,and dynamic computation offloading based on reinforcement learning.The main work content and contributions are as follows.For cooperatively offloading reusable tasks,we study cache-assisted computation offloading in the device-edge and device-edge-cloud networks,respectively.In the device-edge network,we propose a two-stage collaborative offloading algorithm to fully utilize the computing and caching resources in the edge and minimize the overall energy cost.We decompose the proposed minimization problem into two different multidimensional knapsack problems in two stages,design two items,energy data-size ratio and maximum energy savings,to implement the greedy algorithms,and use the device with the highest RE to upload the task data.The simulation result shows that compared to several baseline algorithms,our proposal not only achieves the lowest total energy consumption but also avoids the rapid energy depletion on the low-energy device.In the device-edge-cloud network,considering the reusability of cloudside service data,unknown and variable task popularity,time-varying channel conditions,and unknown computing resources,we propose a novel Online Joint Optimization Framework(OJOF)with decentralized offloading strategy and popularity-aware caching strategy to minimize the total task execution delay.In the OJOF,we design an "alternate-decision,parallel-execution" mechanism to adjust the offloading and caching processes;we build the online offloading problem on each device as a contextual multi-armed bandit,and propose an iLinUCB-based online offloading algorithm to implement autonomous decision-making;we build the online caching problem on the edge as a knapsack problem related to task popularity,and propose a two-level change point detection based online caching algorithm to estimate the task popularity and make caching strategy through dynamic programming.The simulation result shows that the performance of our proposal is the closest to the performance of the algorithm with perfect information.For offloading reusable tasks in a heterogeneous environment,we study the coalitional game based cooperative computation offloading to implement the optimal offloading strategy and data transmission strategy through incentivedriven device cooperation under the assumption of device selfishness.Specifically,considering that the devices can cooperatively transmit data of a reusable task according to different data ratios,we first formulate two offloading and data transmitting joint optimization problems with the objective of minimizing the task execution cost for single-task and multi-task models,respectively;secondly,we formalize the cooperative offloading process of a reusable task into a coalitional game,propose a Coalitional Game based Cooperative Offloading(CGCO)algorithm for the single-task model,and prove the CGCO can achieve a Nash-stable and optimal solution;thirdly,we expand the CGCO into a CGCOM algorithm for the multiple-task model with jointly applying a two-stage flow shop scheduling approach which helps to obtain an optimal task schedule,and prove the CGCO-M can also achieve a Nash-stable solution.The simulation result shows that the performance of CGCO is optimal,and the performance of CGCO-M is the closest to optimal.For offloading reusable tasks to multiple edge servers in a dynamic environment,we study the reinforcement learning based dynamic computation offloading to implement the long-term optimal offloading strategy and data transmission strategy in the case of limited prior information.Specifically,considering the uncertain task arrival,different edge server assignments,changing channel waiting,and various edge server waiting,we first formulate a sequential decision problem with discrete and continuous variables to minimize the long-term average total task execution cost;secondly,we decompose the minimization problem into two subproblems,a tractable linear programming problem for optimizing the short-term continuous data transmission strategy and a Markov decision process only with discrete actions for optimizing the long-term discrete offloading strategy;thirdly,we develop an Average Reward Proximal Policy Optimization(ARPPO)without discount factor to solve the Markov decision problem,and combine the linear programming solver and ARPPO into our hybrid-ARPPO algorithm with an action-mask layer.The simulation result shows the effectiveness of our hybrid-ARPPO in different system scales and task arrival patterns.