基于深度强化学习的模型并行化研究

Research on parallel computing based on deep reinforcement learning

深度学习并行化在加速模型训练、提高预测精度等方面具有重要作用。文章从数据并行、梯度累积算法两方面对深度交通时间预测模型(Travel Time Estimation Based on Deep Neural Networks,Deep TTE)进行了并行优化。实验以成都出租车数据作为数据集进行了并行模式的训练与评估,实验效果相较于Deep TTE,收敛速度有明显提升,RMSE降低了50.12%,MAPE降低了62.3%,MAE降低了56.02%。

Parallelization of deep learning plays an important role in accelerating model training and improving prediction accuracy.This paper optimizes the Travel Time Estimation Based on Deep Neural Networks(Deep TTE)from two aspects of data parallel and gradient accumulation algorithm.In the experiment,Chengdu taxi data is used as the data set to train and evaluate the parallel mode.Compared with Deep TTE,the experimental effect has significantly improved the convergence speed,with RMSE reduced by 50.12%,MAPE reduced by 62.3%,and MAE reduced by 56.02%.