参数并行:一种基于群启发式算法的机器学习参数寻优方法

Parallel Parameters:An Optimization Method of Machine Learning Parameters Based on Swarm Heuristic Algorithm

为了提高机器学习算法超参数寻优效率,提出了一种基于参数并行机制的机器学习参数寻优方法。该方法通过群启发式算法来进行机器学习算法的参数寻优,将种群转换为Spark平台特有的弹性分布式数据集,针对参数寻优耗时特点并行计算种群中个体适应度。选取随机森林和遗传算法作为实验算法,设计了多组实验对所提出的学习训练方法进行验证。实验结果表明:该方法的参数寻优能力和效率都优于主流的网格搜索算法;在20万条以下的小数据量下,与基于数据并行机制的机器学习参数寻优方法相比,该方法运行时间最多能够减少69.5%,并具有良好的可扩展性。

In order to improve the super parameter optimization efficiency of machine learning algorithm,a machine learning parameter optimization method based on parameter parallel mechanism was proposed.The group heuristic algorithm was used to optimize the parameters of the machine learning algorithm,the population was transformed into the unique elastic distributed data set of Spark platform,and the individual fitness in the population was calculated in parallel according to the time-consuming characteristics of parameter optimization.Random forest and genetic algorithm were selected as experimental algorithms,and several groups of experiments were designed to verify the proposed learning and training method.The experimental results show that the parameter optimization ability and efficiency of this method are better than the mainstream grid search algorithm.Compared with the machine learning parameter optimization method based on data parallel mechanism,the running time of this method can be reduced by 69.5%at most,and has good scalability.