Multi-objective Optimization Method And Application Of Distributed Learning Based On Finite State Automaton

Distributed learning is a highly efficient method of training models using large datasets;however,its execution process leads to high energy consumption due to data transmission,computation,and communication between nodes in a distributed environment.As energy consumption increases,so does carbon emission.Simply reducing data size to mitigate carbon emissions declines model accuracy.Therefore,distributed learning faces a significant challenge to balance the tradeoff between model accuracy and carbon emissions control.Reducing the carbon footprint of distributed learning,while ensuring model accuracy,is an important problem that needs to be addressed.This paper proposes a multi-objective optimization method for distributed learning based on finite-state automata,featuring the following key contributions:(1)The paper uncovers the problems of high energy consumption and carbon emissions in actual distributed learning execution process by examining its computational workload and training process.The paper highlights the importance of optimizing carbon emissions while ensuring model accuracy based on related research findings on distributed learning optimization.(2)An innovative Carbon-aware Resource Scheduling Framework(CRSF)for distributed resource scheduling that is highly sensitive to carbon emissions is proposed to resolve the conflict between model accuracy and carbon emissions.This framework optimizes the energy consumption and carbon emissions of distributed learning at both the global scheduling and local optimization levels.The global scheduling layer utilizes the finite-state automata to allocate computational resources that minimize carbon emissions in distributed learning overall.In contrast,the local optimization layer zeroes in on optimizing the distributed learning model training process,combining incremental learning and forgetting learning to alleviate carbon emissions while maintaining model accuracy.(3)The paper’s proposed framework is verified through multiple indicators,including model accuracy,training carbon emissions,training energy consumption,and renewable energy utilization rate.The experimental results demonstrate that the proposed method has an average reduction of 79.64%in energy consumption and 73.24%in carbon emissions while preserving the accuracy of the distributed learning model in various testing sets.