Continual Learning for Adaptable Car-Following in Dynamic Traffic Environments

TL;DR

The paper tackles the challenge of distributional shift in car-following by enabling continual learning through Elastic Weight Consolidation and Memory Aware Synapses, preventing catastrophic forgetting during incremental updates. An LSTM-based baseline is extended with CL losses and evaluated on Waymo and Lyft data across three task sets defined by mean following-speed, showing superior learning and safety performance. The CL-EWS and CL-MAS approaches achieve lower mean-squared errors and maintain a 0% collision rate across all traffic scenarios, highlighting robustness and safety in dynamic environments. This work advances autonomous driving by delivering adaptable, safer car-following behavior through principled continual-learning techniques.

Abstract

The continual evolution of autonomous driving technology requires car-following models that can adapt to diverse and dynamic traffic environments. Traditional learning-based models often suffer from performance degradation when encountering unseen traffic patterns due to a lack of continual learning capabilities. This paper proposes a novel car-following model based on continual learning that addresses this limitation. Our framework incorporates Elastic Weight Consolidation (EWC) and Memory Aware Synapses (MAS) techniques to mitigate catastrophic forgetting and enable the model to learn incrementally from new traffic data streams. We evaluate the performance of the proposed model on the Waymo and Lyft datasets which encompass various traffic scenarios. The results demonstrate that the continual learning techniques significantly outperform the baseline model, achieving 0\% collision rates across all traffic conditions. This research contributes to the advancement of autonomous driving technology by fostering the development of more robust and adaptable car-following models.

Figures (5)

• Figure 1: The densities of the training data (green) and prediction data (orange) reveal a distributional shift. This mismatch could lead to performance degradation on the prediction task.
• Figure 2: The impact of Elastic Weight Consolidation (EWC) on loss during continual learning. It illustrates how EWC helps mitigate the trade-off between learning new tasks and forgetting old tasks.
• Figure 3: The distribution of mean FV speed across each event for all task sets.
• Figure 4: Comparison of training flows for continual learning car-following models: Baseline vs. EWC vs. MAS.
• Figure 5: Comparisons of trajectories performed by different models after training on all three task sets.