Continual Learning of Microscopic Traffic Models Using Neural Networks

In a mixed-autonomy traffic scenario, where human drivers and autonomous vehicles share the streets, self-driving cars need to be able to predict in a robust manner the behaviour of human-driven vehicles, in order to guarantee a safe and smooth driving experience. Although traffic theory provides several models of human drivers, these models are often parameterized by few parameters which can limit their performance in modeling complex behaviors. The lack of sufficient model capacity and the behavioral shifts in human driving reduces the usefulness of these methods in real-life situations. Based on recent advances in trust region optimization, the authors present a new method for data-efficient continual learning, that allows to incrementally train a high-performance driver model, while avoiding the effects of catastrophic forgetting. The proposed approach focuses on keeping output distributions of previous tasks stable during training in new scenarios, by using explicit constraints in the optimization problem based on Kullback–Leibler divergence. As a result, the authors observe minimal loss of performance in previous tasks, while increasing the generalization capabilities of the learned representations. The authors evaluate the performance of the proposed method in traffic modeling tasks, including mandatory lane change and acceleration tasks. Vehicle trajectory data from Next Generation Simulation (NGSIM) is used for training and validation of the models. Results show state-of-the-art performance in the presence of scenarios with small amounts of data.