Source-linked AI summary
TrajGAIL: Generating Urban Vehicle Trajectories using Generative Adversarial Imitation Learning
Seongjin Choi, Jiwon Kim, Hwasoo Yeo
TL;DR
Existing discriminative approaches are limited for generating full urban vehicle trajectories from sparse observations, motivating models that learn trajectory distributions for synthetic data generation. TrajGAIL formulates trajectory generation as imitation learning in a POMDP and uses an adversarial discriminator for training. The model generates realistic trajectories, matches route distributions, and outperforms RNN in most test cases, while remaining limited by omitted traffic and interaction variables.
Problem
Existing discriminative models have limitations in generating full trajectories from sparse samples, while synthetic trajectories are needed to address urban mobility data sparsity and privacy.
Method
TrajGAIL formulates driver movement decisions as a POMDP and learns the demonstrated decision process through generative adversarial imitation learning.
Results
TrajGAIL generates realistic trajectories capturing sequence patterns and route distributions, and outperforms RNN in most test cases, especially for route-distribution matching.
Takeaways & Limitations
TrajGAIL provides a data-driven alternative for synthetic route-distribution generation without requiring OD identification or computationally expensive route-set enumeration.
Takeaways & Limitations
The current model omits traffic conditions, vehicle interactions, and additional predictors such as origin-destination and temporal information.
Abstract
from arXiv · showhide
Recently, an abundant amount of urban vehicle trajectory data has been collected in road networks. Many studies have used machine learning algorithms to analyze patterns in vehicle trajectories to predict location sequences of individual travelers. Unlike the previous studies that used a discriminative modeling approach, this research suggests a generative modeling approach to learn the underlying distributions of urban vehicle trajectory data. A generative model for urban vehicle trajectories can better generalize from training data by learning the underlying distribution of the training data and, thus, produce synthetic vehicle trajectories similar to real vehicle trajectories with limited observations. Synthetic trajectories can provide solutions to data sparsity or data privacy issues in using location data. This research proposesTrajGAIL, a generative adversarial imitation learning framework for the urban vehicle trajectory generation. In TrajGAIL, learning location sequences in observed trajectories is formulated as an imitation learning problem in a partially observable Markov decision process. The model is trained by the generative adversarial framework, which uses the reward function from the adversarial discriminator. The model is tested with both simulation and real-world datasets, and the results show that the proposed model obtained significant performance gains compared to existing models in sequence modeling.
Highlights
TrajGAIL is presented as a framework for urban vehicle trajectory generation using generative adversarial imitation learning, with evaluation targeting both trajectory-level and dataset-level similarity.
- TrajGAIL applies generative adversarial imitation learning to urban vehicle trajectory generation.
- The paper evaluates both trajectory-level similarity and distributional similarity between generated and real trajectory datasets.
1. Introduction
Urban vehicle trajectory data offer detailed mobility information, but existing datasets are incomplete and common discriminative models do not capture the underlying data distribution. The paper proposes TrajGAIL, which combines generative adversarial imitation learning with a partially observable Markov decision process to generate synthetic trajectories.
- Urban trajectory data provide information about aggregated traffic flows and individual travel behavior in road networks.
- Generative models learn the underlying joint distribution and can produce synthetic data by sampling from it.This contrasts with discriminative models, which learn decision boundaries or conditional predictions without modeling the underlying distribution.
- Synthetic trajectory generation addresses data sparsity and privacy-related challenges in urban vehicle trajectory analysis.Available datasets cover only a portion of vehicles, motivating trajectories that more fully represent network traffic dynamics and performance.
- Existing next-location prediction models can generate trajectories sequentially, but most use discriminative modeling and predict one or two next locations.The introduction identifies these models as limited because they do not learn the underlying trajectory distribution.
- TrajGAIL applies imitation learning to generate urban vehicle trajectories from expert demonstrations and learned sequential decision-making behavior.The framework uses the generative adversarial imitation learning approach to model trajectory generation.
- TrajGAIL incorporates a partially observable Markov decision process because vehicle location choices depend on previous as well as current positions.The POMDP maps sequences of location observations into a latent state, supporting more generalization than a state representation based only on current position.
2. Methodology
TrajGAIL models urban vehicle trajectory generation as sequential decision making under partial observability, combining POMDP state representations with GAIL-based imitation learning. Its recurrent components incorporate visit history, while adversarial rewards train a generator to produce realistic link sequences.
- Problem formulation: TrajGAIL defines trajectory generation as learning the probability distribution of discrete road-link sequences.The location sequence is represented using visited link IDs, and the joint distribution is modeled sequentially.
- Problem formulation: The road-network MDP treats links as states, link transitions as actions, and the policy as a driver’s route-choice behavior.The transition model deterministically maps a current link and movement choice to the next link.
- Problem formulation: Because link choices depend on previous locations and unobserved trip or traffic information, the model replaces the MDP assumption with a POMDP.Including all relevant information directly in the state would create an extremely large, potentially intractable state space.
- Problem formulation: TrajGAIL represents observations as link IDs plus Start and End tokens, and uses Straight, Left, Right, and Terminate actions for the grid network.A masking approach can extend the action representation to networks with more diverse intersection structures.
- TrajGAIL framework: POMDP history embedding addresses looping trajectories by incorporating visited-link sequences, while GAIL supplies discriminator-based rewards for future decision making.Standard GAIL produced excessively long trajectories with repeated loops; TrajGAIL uses an RNN embedding layer to represent visit history.
- TrajGAIL framework: The discriminator distinguishes real from generated trajectories, and separate recurrent embeddings let the policy, value, and discriminator modules interpret observation sequences differently.The discriminator processes observation sequences and the final action, while separate objectives update the related networks.
3. Performance Evaluation
TrajGAIL was evaluated on simulated AIMSUN trajectories and real Gangnam taxi trajectories using trajectory-level and dataset-level similarity measures. It generally remained robust as trajectory complexity increased, outperforming other models particularly in matching route distributions, although all models degraded on sparse real-world data.
- Datasets: TrajGAIL was evaluated on AIMSUN simulation data and passenger-level sub-trajectories extracted from Gangnam taxi data.The evaluation used both virtual and real-world trajectory datasets.
- Trajectory-level Evaluation: All models performed well on Single-OD datasets, with four models exceeding 0.99 on both BLEU and METEOR across demand types.MaxEnt(SVF) was the exception, reaching 0.9627 under Proportional demand.
- Trajectory-level Evaluation: As dataset complexity increased, RNN and TrajGAIL showed only slight score decreases and achieved BLEU = 1.0 and METEOR = 1.0 in most cases.On Gangnam, TrajGAIL averaged 0.9974 for both BLEU and METEOR, compared with RNN averages of 0.9726 and 0.9899.
- Dataset-level Evaluation: Route-distribution evaluation used Jensen-Shannon distance, where lower values indicate better statistical similarity between generated and real datasets.Route distribution is difficult to match because it depends on matching several other trajectory variables.
- Dataset-level Evaluation: TrajGAIL performed noticeably better on complex Multi-OD datasets because it combines sequential embeddings with discriminator-derived rewards.The discriminator reward provides additional guidance for generating trajectories that match real observations.
- Dataset-level Evaluation: On the Gangnam dataset, all models had d_JS above 0.4, but TrajGAIL generated only 181 unknown trajectories among 20,000 samples.The authors attribute the degradation to sparse rare routes in the real-world dataset; the other models generated more than 2,000 unknown trajectories.
- Discussion: TrajGAIL and RNN were the most suitable models overall, with TrajGAIL outperforming RNN in most test cases, especially for route-distribution matching.The discussion attributes TrajGAIL’s advantage to training signals supplied by its modeling architecture.
4. Conclusion
TrajGAIL models urban vehicle trajectory generation as imitation learning in a partially observable decision process and generates synthetic trajectories resembling real observations. Evaluations compare trajectory-level and dataset-level similarity across datasets and demand patterns, while proposed extensions add contextual information.
- Core approach: TrajGAIL formulates drivers’ movement decisions as a partially observable Markov decision process and learns the demonstrated decision process through generative adversarial imitation learning.The framework generates synthetic trajectory data similar to real-world trajectory observations.
- Methodological contribution: TrajGAIL combines POMDP modeling with RNN embeddings to encode historical sequences while preserving the Markov assumption required by the imitation-learning framework.This addresses the inability of standard GAIL and traditional IRL methods to use previous locations when selecting the next location.
- Evaluation: The evaluation measures both trajectory-level similarity using BLEU and METEOR and dataset-level statistical similarity across datasets with varied traffic demand patterns.Generated trajectories are compared against three baseline models from previous studies.
- Extensions: Future extensions could incorporate traffic conditions, origin and destination information, and temporal variables such as time of day and day of week.The current study focuses on theoretical aspects and structural differences between TrajGAIL and other models.
- Scope boundary: TrajGAIL does not currently account for traffic conditions or interactions with other vehicles and therefore cannot serve as a route-choice model for dynamic traffic assignment or traffic simulation.The paper identifies incorporating additional information as a future research direction.
CRediT authorship contribution statement
The contribution statement assigns conceptualization, methodology, software, validation, formal analysis, writing, supervision, data curation, resources, funding acquisition, and project administration across the authors.
- Seongjin Choi handled conceptualization, methodology, software, validation, formal analysis, and the original draft.
- Jiwon Kim handled conceptualization, methodology, formal analysis, supervision, and review and editing.
- Hwasoo Yeo handled data curation, resources, supervision, funding acquisition, project administration, and review and editing.