Enhancing Social Motion Prediction Using Attention Mechanisms and Hierarchical Structures

Botao Dong, Yumo Ji, Yongze Miao

2024

Abstract

With the continuous deepening of artificial intelligence applications in multi-agent systems, the accuracy and efficiency of motion prediction have become key research challenges. This paper proposes a novel neural network model based on Kolmogorov-Arnold Networks (KANs) aimed at enhancing the generalization ability and prediction accuracy of models in multi-agent motion prediction tasks. The study first analyses the limitations of existing behavioural cloning methods and Generative Adversarial Imitation Learning (GAIL) in handling complex dynamic interactions and nonlinear feature data. To address these issues, this paper introduces KANs, a model that replaces the weight parameters in traditional multi-layer perceptron with learnable univariate spline functions, thereby enhancing the model's nonlinear feature extraction capability and adaptability. In the experiments, this paper adopts the Wusi dataset proposed by Zhu et al. in 2024, which contains historical motion sequences of multiple participants. The model designed in this study combines Transformer encoders and decoders, along with KANs, to process local and global features and generate motion predictions for all participants in future time periods. Through feature fusion nodes and multi-level strategy networks, the model can generate more natural and accurate motion sequences. The experimental results show that compared with traditional Transformer-based models, the model in this paper has significantly improved prediction accuracy and training efficiency. Moreover, the model demonstrates better generalization ability on unseen complex patterns, providing new perspectives and methods for the practical application of multi-agent systems.

Paper Citation

in Harvard Style

Dong B., Ji Y. and Miao Y. (2024). Enhancing Social Motion Prediction Using Attention Mechanisms and Hierarchical Structures. In Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management - Volume 1: MLSCM; ISBN 978-989-758-738-2, SciTePress, pages 185-189. DOI: 10.5220/0013253300004558

in Bibtex Style

@conference{mlscm24,
author={Botao Dong and Yumo Ji and Yongze Miao},
title={Enhancing Social Motion Prediction Using Attention Mechanisms and Hierarchical Structures},
booktitle={Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management - Volume 1: MLSCM},
year={2024},
pages={185-189},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0013253300004558},
isbn={978-989-758-738-2},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management - Volume 1: MLSCM
TI - Enhancing Social Motion Prediction Using Attention Mechanisms and Hierarchical Structures
SN - 978-989-758-738-2
AU - Dong B.
AU - Ji Y.
AU - Miao Y.
PY - 2024
SP - 185
EP - 189
DO - 10.5220/0013253300004558
PB - SciTePress