Research And Implementation Of Distributed Environment Trajectory Prediction Algorithm

Trajectory prediction technology is widely used in reality.With the arrival of the era of big data,using large number of distributed data to train trajectory prediction model has become the key research direction of trajectory prediction technology.At present,we generally use distributed neural network framework to train trajectory prediction model.However,there exists some problems in the current distributed neural network framework,such as low communication efficiency,low model accuracy and so on,which results in the lack of ideal training speed and accuracy of model.In order to solve the problem of low convergence efficiency and accuracy of trajectory prediction model training in a distributed environment,this thesis focus on the performance bottleneck of trajectory prediction model framework,trying to improve trajectory prediction model's convergence efficiency and accuracy.The key work and innovations are as follows:(1)An adaptive gradient compression algorithm is proposed.During the training phase of neural network in distributed environment,neural network needs to transmit a large number of intermediate gradients between computing nodes.With the improvement of the efficiency of stand-alone optimization of computing nodes,network transmission has become the main factor affecting the training efficiency in distributed neural networks.Therefore,it is necessary to compress the transmission gradient to shorten the training time.However,the traditional gradient compression algorithm usually can't effectively adjust the compression ratio dynamically,so it is impossible to balance the two dimensions of model accuracy and training efficiency at the same time.Therefore,this thesis dynamically perceiving the process of model convergence according to the characteristics of neural network model training,and adjusts the gradient compression ratio to improve the convergence speed of the model as much as possible with the minimum accuracy reduction cost.Experiments show that the algorithm is feasible in the environment where the network communication cost is high.(2)A self-organizing packet dynamic hybrid synchronization algorithm is proposed.The communication pace between different computing nodes directly affects the accuracy and speed of model training.Currently,there are three main communication algorithms:synchronous communication,asynchronous communication and hybrid communication,in which synchronous communication algorithm has the highest model accuracy and asynchronous communication algorithm has the highest model training speed.On the basis of ensuring the accuracy,how to make efficient use of the computing performance of all computing nodes is the key research direction of communication algorithms.Based on the traditional communication algorithm,this thesis proposes a self-organizing packet dynamic hybrid synchronization algorithm,which dynamically adjusts the communication scheme according to the cluster context to improve the convergence speed of the model.(3)Build a high-performance distributed trajectory prediction model platform.On the basis of the above two algorithms,a high-performance distributed trajectory prediction model platform is designed,which integrates computing node management,task management,model prediction as well as extra functions,and it interacts with platform users in a visual way.