Conformalized Adaptive Forecasting of Heterogeneous Trajectories
Authors: Yanfei Zhou, Lars Lindemann, Matteo Sesia
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This section demonstrates the empirical performance of our method. We focus on applying CAFHT with multiplicative scores, based on the ACI algorithm, and tuning the learning rate through data splitting. Additional results pertaining to different implementations of CAFHT are in Appendix A5. In all experiments, the candidate values for the ACI learning rate parameter γ range from 0.001 to 0.1 at increments of 0.01, and from 0.2 to 0.9 at increments of 0.1. |
| Researcher Affiliation | Academia | 1Department of Data Sciences and Operations, University of Southern California, Los Angeles, CA, USA 2Department of Computer Science, University of Southern California, Los Angeles, CA, USA. |
| Pseudocode | Yes | Algorithm 1 CAFHT |
| Open Source Code | Yes | Software implementing the algorithms and data experiments are available online at https://github.com/ Fiona Z3696/CAFHT.git. |
| Open Datasets | Yes | We now apply the three methods to forecast pedestrian trajectories generated from the ORCA simulator (Van den Berg et al., 2008), which follow nonlinear dynamics and are intrinsically harder to predict than the synthetic trajectories discussed before. The data include 2-dimensional position measurements for 1,291 pedestrians, tracked over T = 20 time steps. |
| Dataset Splits | Yes | In each case, 75% of the trajectories are used for training and the remaining 25% for calibration. Our method utilizes 50% of the calibration trajectories to select the ACI learning rate γ. |
| Hardware Specification | No | The authors thank anonymous referees for helpful comments, and the Center for Advanced Research Computing at the University of Southern California for providing computing resources. |
| Software Dependencies | No | For all methods, the underlying forecasting model is a recurrent neural network with 4 stacked LSTM layers followed by a linear layer. The learning rate is set equal to 0.001, for an Adam W optimizer with weight decay 1e-6. |
| Experiment Setup | Yes | For all methods, the underlying forecasting model is a recurrent neural network with 4 stacked LSTM layers followed by a linear layer. The learning rate is set equal to 0.001, for an Adam W optimizer with weight decay 1e-6. The models are trained for a total of 50 epochs, so that the mean squared error loss loss approximately converges. |