自适应多智能体算法优化深度网络的列车智能驾驶

Intelligent train operation based on adaptive multi-agent algorithm optimization for deep network

在利用深度学习实现列车智能驾驶时,神经网络架构和参数的选择过于依赖人工经验,现有梯度下降法在参数优化时易陷入局部最优,且学习任务单一。针对上述问题,提出一种自适应多种群链式多智能体算法(AMPCMA)优化LSTM网络的列车智能驾驶新方法,该方法有机地将计算智能与深度学习结合,能充分挖掘优秀驾驶员数据。其具体实现过程为:首先,从自动化深度学习角度出发,采用遗传算法(GA)优化深度网络结构,克服了其结构难以确定的问题,并在此基础上分粗、细学习2个阶段对整个网络的参数进行优化。在粗学习阶段,采用AMPCMA算法对LSTM参数预置初值,有效地学习多个任务的共性。该算法能在进化过程中动态调整小种群链表规模,具有较好的灵活性和自适应性。接着在细学习阶段,基于上述多任务共性学习所得到的LSTM参数,再用Adam算法分别对单个任务上的参数精细优化,以实现任务的个性学习;其次,有效地设计了多任务之间的信息共享机制,且任务共性和个性学习有机结合,使得整个网络泛化能力强,较好地改善了列车档位、档位操纵时间和列车速度的多任务决策效果;最后,通过仿真实验验证了所提出的AMPCMA-LSTM模型较传统机器学习方法更优越,提高了列车操控与预测精度,并能在多种操控序列下表现出较强的鲁棒性。

In the application of deep learning to realize intelligent train operation,the selection of the neural network architecture and parameters excessively relies on manual experience,while the existing gradient descent method is prone to fall into local optimum in parameter optimization and the learning task is single.In order to address these issues,a new method of intelligent train operation based on an adaptive multi-population chainlike multi-agent algorithm(AMPCMA)was proposed to optimize the Long Short-Term Memory(LSTM)network in this paper.The proposed method can organically combine computational intelligence with deep learning to fully mine excellent driver data.The specific process of implementation is as follows.First,from the perspective of automatic deep learning,the genetic algorithm(GA)was used to optimize the structure of the deep network,which can overcome the problem of difficulty in determining its structure.On this basis,two stages consisting of coarse and fine learning were utilized to optimize the parameters of the whole network.In the coarse learning stage,the initial parameters of LSTM were set with AMPCMA to effectively learn the commonality of multiple tasks.This algorithm can dynamically adjust the scale of the small population chain table during the evolution process,which has good flexibility and adaptability.In the fine learning stage,based on the LSTM parameters obtained from the above multi-task learning,the Adam algorithm was used to finely optimize the respective parameters of a single task to achieve individual learning on the task.Secondly,the information sharing mechanism among multi-tasks was effectively designed,and the commonality and individuality learning of tasks were organically combined,so as to make the generalization ability of the whole network stronger,and further improve the decision-making effect of multi-task about train manipulation gear,gear operating time and train speed.Finally,the proposed AMPCMA-LSTM model was verified to be superior to the traditional machine learning methods through simulation experiments.It can not only improve the accuracy of train manipulation and prediction,but also show strong robustness under various manipulation sequences.