Attention-based Neural Networks for Multi-modal Trajectory Prediction

Emami, Negar; Braun, Torsten (6 May 2022). Attention-based Neural Networks for Multi-modal Trajectory Prediction. In: Bern Data Science Day. University of Bern: Data Science Lab

Preview	Text EmamiNegar_BernDataScience2022.pdf - Supplemental Material Available under License BORIS Standard License. Download (57kB) | Preview
Preview	Other (Poster) BDSD22_Poster_NegarEmami-3.pdf - Presentation Available under License BORIS Standard License. Download (626kB) | Preview

Trajectory prediction is of great importance in wireless and intelligent networks. Accurate forecast of users’ trajectories can provide efficient handover management, continuous network connection, and generally better network quality of service. A trajectory is defined as the sequence of location logs, e.g., GPS coordinates or cellular antenna IDs, over time. We present a trajectory predictor based on Transformers Neural Networks acquiring the self-attention mechanism [1]. Mobile objects’ mobility patterns are influenced by their nearby neighbors. Thus, learning spatio-temporal dependencies among neighbor-trajectory users can help to better predict their trajectories [2]. In this direction, unlike our previously proposed mobility predictor (based on LSTM and CNN) designed for single agents [3], [4], [5], where agents were acting in isolation, we now propose the INteractive TRAnsformers ReinFORCEd (INTRAFORCE) social-aware neural network. We further employ a reinforcement learning agent to design the highest-performance transformers neural architecture based on the multi-modal trajectory scenario. Evaluations show that using the Orange dataset [4], our transformer-based predictor can remarkably increase the accuracy and decrease the training time and computations concerning our models based on LSTM and CNN [4]. Furthermore, on ETH+UCY datasets [6], INTRAFORCE achieves the least Mean Square Error compared to numerous state-of-the-art mechanisms on this popular dataset.

Interest & Impact

Downloads

0 since deposited on 03 May 2023
0 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