基于PBT的自动调参算法应用研究

Research on the Application of PBT-based Automatic Parameter Adjusting Algorithm

超参数的取值对机器学习的模型效果有至关重要的作用。如果根据德尔菲经验法来人工选择超参数,那么需要充足的相关经验、大量的精力以及不可控制的运气成分。深度强化学习的超参数配置最理想的状态是可依据当前所面临的学习状态进行自动调整,并非在多种状态下只使用同一组配置好的超参数。基于此,提出一种基于群体的超参数自动调整算法(PBT)。PBT是参数随机搜索及人工调优这两种使用频率最高的超参数优化方法的结合,其在搜索参数的同时训练并优化参数网络,从而能快速准确地找到最优参数。超参数自动调整的方法可大幅缩短模型构建时间,提升建模效率,提高模型精度。PBT使得试验者能快速为任务选择最佳的超参数集合和模型。

The value of hyperparameters plays a vital role in the model effect of machine learning.If the hyperparameters are manually selected according to the Delphi empirical method,sufficient relevant experience,a lot of energy,and uncontrollable elements of luck are required.The most ideal state of the hyperparameter configuration of deep reinforcement learning is that it can be automatically adjusted following the current learning state,instead of using the same set of configured hyperparameters in multiple states.An algorithm for automatic adjustment of hyperparameters based on population(PBT)is proposed.PBT which is the two most frequently used hyperparameter optimization methods is a combination of manual parameter adjusting and parameter random search.PBT trains and optimizes the parameter network while searching for parameters,so the optimal parameters can be found quickly and accurately.The method of automatically adjusting hyperparameters greatly shortens the model construction time,improves the efficiency of modeling,and improves the accuracy of the model.PBT enables the experimenter to quickly select the best hyperparameter set and model for the task.