Learning human’s car-following behavior needs not only well-designed models but also effective training or calibration methods. Comparing with the vast amount of efforts on car-following modeling in literature, training methods are less studied. This research proposes a training method (BPTS - Backpropagation through Simulation) to reduce the long-term error of neural network-based car-following models, with multiple experimental validations. The training method uses a recurrent framework with simulation to generate long-term predictions for generic car-following models, and use gradient backpropagation to reduce accumulative error. The proposed training method can also calibrate other car-following models besides neural network-based models. In experimental validation, our studies yielded more than 30% error reduction in long-term (20 s) prediction for feed-forward Artificial Neural Network (ANN) and Long short-term memory (LSTM) models, and reduces the error on vehicle position by more than 1.0 meters, at the cost of that short-term (0.2 s) prediction error slightly increases. The proposed training method dramatically reduces the long-term prediction error of neural network-based car-following models.
Backpropagation through Simulation: A Training Method for Neural Network-based Car-following
2019-10-01
2797638 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Fringe pattern optimisation using a backpropagation neural network
British Library Conference Proceedings | 1996
|Optical backpropagation training method and its applications
British Library Conference Proceedings | 2020
|Efficient Training of Neural Networks using Optical Backpropagation with Momentum Factor
British Library Online Contents | 2008
|