Multi-agent Trajectory Prediction with Fuzzy Query Attention
Authors: Nitin Kamra, Hao Zhu, Dweep Kumarbhai Trivedi, Ming Zhang, Yan Liu
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our architecture demonstrates significant performance gains over existing state-of-the-art predictive models in diverse domains such as human crowd trajectories, US freeway traffic, NBA sports data and physics datasets. We also present ablations and augmentations to understand the decision-making process and the source of gains in our model. |
| Researcher Affiliation | Academia | Nitin Kamra Department of Computer Science University of Southern California Los Angeles, CA, USA nkamra@usc.edu Hao Zhu Department of Computer Science, School of EECS Peking University Beijing, China hzhu1998@pku.edu.cn Dweep Trivedi Department of Computer Science University of Southern California Los Angeles, CA, USA dtrivedi@usc.edu Ming Zhang Department of Computer Science, School of EECS Peking University Beijing, China mzhang_cs@pku.edu.cn Yan Liu Department of Computer Science University of Southern California Los Angeles, CA, USA yanliu.cs@usc.edu |
| Pseudocode | No | No explicit pseudocode or algorithm blocks were found. |
| Open Source Code | Yes | Code for implementing FQA can be found at https://github.com/nitinkamra1992/FQA.git |
| Open Datasets | Yes | 1. ETH-UCY [3]: A human crowds dataset with medium interaction density. We sampled about 3400 scenes at random from the dataset and set T = 20 following prior work [1, 8]. 2. Collisions: Synthetic physics data with balls moving on a friction-less 2D plane, fixed circular landmarks and boundary walls. The collisions between balls preserve momentum and energy, while collisions of agents with walls or immobile landmarks only preserve energy but not momentum of moving agents. Contains about 9500 scenes with T = 25. 3. NGsim [5]: US-101 and i-80 freeway traffic data with fast moving vehicles. ... 4. Charges [14]: Physics data with positive and negative charges moving under other charges electric fields and colliding with bounding walls. Contains 3600 scenes with T = 25 involving dense attractive and repulsive interactions. 5. NBA [30]: Sports dataset with basketball player trajectories. We sampled about 7500 scenes with T = 30. This dataset features complex goal-oriented motion heavily dictated by agents intentions. |
| Dataset Splits | Yes | For datasets with no provided splits, we follow a 70 : 15 : 15 split for training, validation and test set scenes. |
| Hardware Specification | No | our simpler row-wise dot-product structure fits easily on a single GPU (12GB) for all datasets |
| Software Dependencies | No | The paper mentions the use of "Adam optimizer [13]" but does not provide specific version numbers for any software dependencies like programming languages, libraries, or frameworks (e.g., PyTorch, TensorFlow, Python). |
| Experiment Setup | Yes | All models are trained with the Adam optimizer [13] with batch size 32 and an initial learning rate of 0.001 decaying multiplicatively by a factor γ = 0.8 every 5 epochs. All models train for at least 50 epochs after which early stopping is enabled with a max patience of 10 epochs on validation set mean-square error and training is terminated at a maximum of 100 epochs. Since we test the models by observing Tobs (kept at 2T / 5 for all datasets) time-steps and make predictions until the remaining time T, we followed a dynamic schedule allowing all models to see the real observations for Ttemp time-steps followed by T - Ttemp of its own last time-step predictions. During training, Ttemp is initialized to T and linearly decayed by 1 every epoch until it becomes equal to Tobs. |