基于人群的神经网络超参数优化的研究

Population-based hyper-parameter optimization of neural networks

神经网络主导了现代机器学习领域,但是他们的训练和成功仍然受到超参数经验选择的敏感影响,例如模型架构,损失函数和优化算法。文中介绍基于人群训练(Population Based Training简称PBT),一种简单的异步优化算法,有效地利用固定的计算预算来优化模型以及其超参数群体的最大化性能。重要的是,PBT发现了一个超参数设置的时间表,而不是遵循一般的次优策略,试图找到一个固定的集合用于整个训练过程。通过对典型分布式超参数训练框架的小步幅修改,文中的方法可以对模型进行稳健可靠的训练。文中展示了PBT在深度强化学习问题上的有效性,通过对一组超参数进行优化,显示了更快的时钟收敛和更高的代理性能。

Neural networks dominate the modern machine learning,but their training and success are still sensitive to the selection of superparametric experience,such as model architecture,loss function and optimization algorithm. This paper introduces population-based Training（ Population-based Training PBT）,a simple asynchronous optimization algorithm,which effectively used to optimize the calculation of fixed budget model and its parameter group to maximize performance. Importantly,the PBT founds a schedule for setting a superparameter,instead of following a general suboptimal strategy,attempts to find a fixed set for the entire training process. By modifying the small step size of a typical distributed superparametric training framework,the method can train the model stably and reliably. This paper demonstrates the effectiveness of PBT in deep reinforcement learning. By optimizing a set of superparameters,it show faster clock convergence and higher final performance of agents.