Abstract
Traditional car-following models in micro-traffic simulations often suffer from significant error propagation, where small inaccuracies in individual vehicle predictions cascade through a platoon over time. This paper introduces the Unidirectional Interconnected LSTM (Int-LSTM), a novel architecture designed to treat a vehicle platoon as a single interconnected system. By integrating a spatial-temporal interconnected structure with a scheduled sampling training technique, the Int-LSTM model reduces trajectory generation errors by approximately 40% compared to standard deep learning approaches.
Introduction
Micro-traffic simulation is a critical tool for evaluating transportation infrastructure and autonomous vehicle systems. Central to these simulations are car-following models, which describe how a following vehicle adjusts its speed and position relative to the lead vehicle. While Long Short-Term Memory (LSTM) networks have shown promise in capturing complex driving behaviors, they are prone to "error propagation." In long-duration simulations, minor errors in the first follower's state lead to larger discrepancies for subsequent vehicles (spatial error) and accumulate over time (temporal error), eventually resulting in unrealistic trajectories. This research addresses these challenges by proposing a model that maintains stability across entire platoons.
Methodology
The proposed Unidirectional Interconnected LSTM (Int-LSTM) addresses error propagation through three primary innovations:
1. Interconnected Spatial-Temporal Structure
Unlike traditional models that treat vehicle pairs in isolation, the Int-LSTM architecture links the hidden states of all vehicles within a platoon. This design acknowledges the unidirectional flow of information in traffic—from the leader to the followers—allowing the model to extract features from the collective dynamics of the group.
2. Scheduled Sampling
A common issue in sequence modeling is the "training-inference mismatch." During standard training (Teacher Forcing), models are fed ground-truth data. However, during inference, they must rely on their own previous, potentially erroneous, predictions. Scheduled sampling bridges this gap by gradually transitioning the model from using ground-truth inputs to its own generated outputs during training, teaching it to self-correct.
3. Platoon-Level Loss Function
The model is trained by minimizing the Mean Squared Error (MSE) across the entire platoon simultaneously. This holistic approach ensures that the learned parameters optimize the performance of the whole system rather than individual components.
Data and Evaluation
The model was trained and evaluated using the Next Generation Simulation (NGSIM) I-80 dataset. To ensure high-quality training data, a reconstructed version of the dataset was used to mitigate sensor noise. The training set comprised 3,752 platoons across various traffic conditions, while the evaluation focused on 73 platoons in congested flow to test the model's robustness during stop-and-go oscillations.
Key Results
The Int-LSTM was benchmarked against the physics-based Intelligent Driver Model (IDM) and a standard LSTM.
- Quantitative Performance: The Int-LSTM achieved a Mean Absolute Error (MAE) of 13.5 meters, a 46% improvement over the standard LSTM (25.1 meters) and a 21% improvement over the IDM.
- Qualitative Stability: Visual analysis confirms that the Int-LSTM maintains high fidelity to ground-truth trajectories even during complex traffic oscillations, whereas traditional LSTMs tend to drift significantly over a 50-second simulation window.
Conclusion and Future Work
The Int-LSTM represents a significant advancement in platoon trajectory generation by addressing the fundamental problem of error propagation. While highly effective for longitudinal car-following, future work will explore extending the model to include lateral movements (lane-changing), intersection dynamics, and explicit safety constraints to prevent vehicle collisions in simulated environments.
Summary updated