Research On Distributed Training Optimization Algorithm For Wireless Sensor Network

Wireless sensor networks can provide computing resources for machine learning.However,with the rapid development of wireless sensor networks in recent years,the rising of training data and increasing complexity of models have severely restricted the original way of stand-alone training with GPU computing.Therefore,using edge computing to execute the process of distributed training has become the key technology to improve the efficiency of machine learning and the utilization of network resources.At present,most distributed machine learning systems adopt an architecture named parameter server to achieve data parallelization.However,if the actual edge computing device is used as the worker of distributed training,how to reach the trade-off between the quality of task completion and the execution efficiency of heterogeneous tasks will become a critical issue which requires an urgent settlement.To address the above problems,this paper proposes a heterogeneous task scheduling model in a distributed machine learning system based on the parameter server architecture.The construction of this model mainly relies on the analysis of the calculation process,communication process and the factor of energy consumption involved.In the aforementioned processes,wireless network interference and parameter server decision will all have an impact on the design of heterogeneous task scheduling strategy based on training performance optimization goals.To solve this problem,this thesis introduces relevant conclusions about the distributed machine learning system,conducts in-depth analysis on the convergence efficiency of the distributed machine learning system,and finally determines the influencing factors and corresponding constraints of the optimization goal of distributed training performance so as to complete the formal description of the problem.In addition,the quality of training data will greatly affect the optimization of distributed training performance,which depends on the data collection strategy of the wireless sensor network.Charging the sink node requires a certain cost.Towards balancing the relationship between the charging cost and the optimization goal of distributed training performance,it is necessary to consider introducing a charging reward factor to adjust the energy transmission power,and to stimulate the wireless sensor network to further optimize the training data collection process.In this thesis,the system utility is defined as the overall utility produced by the charging reward factor and training performance under the Pareto efficiency model,and the objective equation is reformulated without changing the relevant constraints and the objective function form.Through equivalent conversion of the objective function and its constraints,etc.,the similarity connection between the objective function and the 0-1 knapsack problem is established from a theoretical perspective,which proves the NP hardness of the proposed problem.To tackle the above problem,this thesis proposes a tabu search algorithm based on greedy strategy.Then,a better initial solution is obtained through the greedy strategy,which reduces the instability of the tabu search algorithm,and finally obtains the approximately optimal solution of the problem.In addition,the algorithm is evaluated through simulation experiments to analyze the execution efficiency and the impact of various factors on system utility.Firstly,five performance indicators under five algorithms are compared at the cluster level and the training task level,including the system utility of the cluster,the ratio of the number of nodes performing training tasks to the number of edge nodes,and the energy consumption produced by task execution in the maximum energy consumption of the distributed cluster.In addition,experiments are conducted to further study the influence of wireless network interference factors and the decision factors of parameter server on the system utility under different conditions.The results demonstrate that the proposed algorithm achieves the superior execution performance compared with existing algorithms.