Reinforcement Learning-designed LSTM for Trajectory and Traffic Flow Prediction

Karimzadeh, Mostafa; Aebi, Ryan; M. de Souza, Allan; Zhao, Zhongliang; Braun, Torsten; Sargento, Susana; Villas, Leandro (5 May 2021). Reinforcement Learning-designed LSTM for Trajectory and Traffic Flow Prediction. In: IEEE Wireless Communications and Networking Conference (pp. 1-6). IEEE 10.1109/WCNC49053.2021.9417511

	Text INFOCOM_Mostafa.pdf - Published Version Restricted to registered users only Available under License Publisher holds Copyright. Download (974kB) | Request a copy
	Text Reinforcement Learning-designed LSTM for Trajectory and Traffic Flow Prediction.pdf - Published Version Restricted to registered users only Available under License Publisher holds Copyright. Download (1MB) | Request a copy

Trajectory and traffic flow prediction will play an essential role in Intelligent Transportation Systems (ITS) to enable a whole new set of applications ranging from traffic management to infotainment applications. In this scenario, deep learning approaches such as Recurrent Neural Networks (RNN) and its variant Long Short Term Memory (LSTM) are excellent alternatives due to their ability to learn spatiotemporal dependencies. However, these neural networks tend to be over-complex and hard to design due to the broad set of hyper-parameters. We propose an automated framework to predict future trajectories and traffic flows in urban areas without human interventions. We employ Reinforcement Learning (RL) and Transfer Learning (TL) to generate high-performance LSTM predictors, which is referred as RL-LSTM. In addition, we introduce HERITOR (High ordEr tRaffIc convoluTiOn Rl-lstm), a novel deep learning algorithm for traffic flow prediction. Specifically, HERITOR attempts to capture pure spatiotemporal features of urban traffic. The extracted features are fed into the RL-LSTM to realize a high performance LSTM for traffic flow prediction. We examine the proposed trajectory and traffic flow predictors on two realworld, large-scale datasets and observe consistent improvements of 15% - 25% over the state-of-the-art.

Interest & Impact

Downloads

118 since deposited on 31 Jul 2019
2 in the past 12 months

Citations

5 Citations in Web of Science ®
6 Citations in Scopus

Search

in Google Scholar™

Services

Actions (login required)

Edit item

Actions (login required)

Edit item