Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
DutyTTE: Deciphering Uncertainty in Origin-Destination Travel Time Estimation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DutyTTE, which uses policy-gradient deep RL (SCST with DTW/LCS rewards) to improve OD path prediction and a mixture-of-experts uncertainty module with MIS loss to output travel-time confidence intervals.
Problem
Quantifying origin-destination travel time uncertainty requires predicting a path matching the ground truth and modeling segment-level travel time uncertainty under varying conditions; existing autoregressive path prediction suffers error accumulation.
Approach
Optimize path generation with a sequence-level reinforcement objective (DTW/LCS-based reward with self-critical baseline) instead of per-step likelihood, and use a noisy top-k mixture-of-experts over segment travel-time distributions to capture context-dependent uncertainty, supervised by mean interval score loss.
Key results
- On Chengdu, DutyTTE reports RMSE 278.22 s, MAE 195.70 s and MAPE 23.55%, versus best baseline T-WDR-MIS at 306.89 s / 221.52 s / 27.92%.
- Uncertainty PICP reaches 91.02% (Chengdu) and 91.67% (Xi'an), above baselines around 84-85% for T-WDR-MIS, at 90% target confidence.
- Path prediction LCS improves to 14.52 (Chengdu) and 14.64 (Xi'an) versus GDP's 13.51 and 13.91.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DutyTTE instead of T-WDR-MIS | Travel time uncertainty quantification · Chengdu | PICP (%): 84.57 -> 91.02 | Helped |
| Use DutyTTE instead of T-WDR-MIS | Travel time uncertainty quantification · Xi'an | PICP (%): 84.93 -> 91.67 | Helped |
| Use DutyTTE instead of T-WDR-MIS | Travel time uncertainty quantification · Chengdu | IW (s): 829.39 -> 763.65 | Helped |
| Use DutyTTE instead of T-WDR-MIS | Travel time uncertainty quantification · Xi'an | IW (s): 840.28 -> 774.83 | Helped |
| Use DutyTTE instead of T-WDR | Travel time estimation · Chengdu | MAPE (%): 27.92 -> 23.55 | Helped |
| Use DutyTTE instead of T-WDR | Travel time estimation · Xi'an | MAPE (%): 26.81 -> 22.95 | Helped |
| Use DutyTTE instead of GDP | Path prediction · Xi'an | LCS: 13.91 -> 14.64 | Helped |
| Use DutyTTE instead of GDP | Path prediction · Xi'an | DTW: 0.024 -> 0.021 | Helped |
| add MoE-guided uncertainty quantification (ablation) | DutyTTE · Chengdu | PICP: 88.63 -> 91.02 | Helped |
| add MoE-guided uncertainty quantification (ablation) | DutyTTE · Chengdu | IW: 781.52 -> 763.65 | Helped |
| add policy loss to path prediction (ablation) | DutyTTE · Chengdu | RMSE: 302.93 -> 278.22 | Helped |
| add policy loss to path prediction (ablation) | DutyTTE · Chengdu | MAPE: 27.28 -> 23.55 | Helped |
Limitations
- Efficiency is comparable to a transformer baseline but slower than simple DA/MWSL methods during training; DRL training adds cost.
- Evaluated only on two taxi datasets from a single provider (Didi) in two Chinese cities.
Experimental setup
Datasets: Didi Chengdu taxi trajectories Didi Xi'an taxi trajectories
Benchmarks: Didi Chengdu Didi Xi'an
Metrics: RMSE MAE MAPE PICP IW LCS DTW
Baselines: T-WDR-MIS MWSL-TTE DeepOD-MIS DOT-dropout DOT-MIS TEMP GBDT ST-NN Dijkstra's algorithm NMLR Key Segment Transformer
Builds on: Self Critical Sequence Training Mean Interval Score loss wide-deep-recurrent Node2Vec noisy top-k gating mixture of experts
Closest papers by meaning
- C2F-TP: A Coarse-to-Fine Denoising Framework for Uncertainty-Aware Trajectory Prediction (AAAI 2025 · similarity 0.82)
- BO4Mob: Bayesian Optimization Benchmarks for High-Dimensional Urban Mobility Problem (NeurIPS 2025 · similarity 0.82)
- Proper Scoring Rules for Agentic Uncertainty Quantification (NeurIPS 2026 · similarity 0.81)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.81)
- CUQDS: Conformal Uncertainty Quantification Under Distribution Shift for Trajectory Prediction (AAAI 2025 · similarity 0.81)
- Active Sequential Posterior Estimation for Sample-Efficient Simulation-Based Inference (NeurIPS 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.