Next Position Prediction using LSTM Neural Networks
John Violos, Stylianos Tsanakas, Maro Androutsopoulou, Georgios Palaiokrassas, Theodora A. Varvarigou · 2020
Movement data is a valuable source of information in the context of many applications. Deep Learning (DL) provides useful methods for the modeling and the knowledge extraction from them. An adaptation of a DL approach to time series analysis can improve significantly the prediction accuracy with the cost of increasing the training time and consequently the resource demands. In this paper we propose the use of Artificial Neural Networks (ANN) with LSTM layers for the next position prediction of moving objects using a genetic algorithm and a transfer learning method. The genetic algorithm makes a smart search in the the hypothesis space to estimate a close to optimal ANN architecture and the model leverages a transfer learning method to exploit a repository of well trained ANN models in order to speed up the training process. We evaluated our proposed model with real data from the trajectory of various vessels and compared with a state of the art trajectory prediction method and two well-known AutoML meta models. The experimental results showed that the proposed model has better accuracy than other models and the transfer learning process decreases dramatically the training time. The results encourage us for the applicability of our method.