A Novel Gated Recurrent Unit Network Based on SVM and Moth-Flame Optimization Algorithm for Behavior Decision-Making of Autonomous Vehicles

Abstract

The behavior decision-making algorithm plays an important role in ensuring the safe driving of autonomous vehicles. However, existing behavior decision-making methods lack the capability to cope with future motion uncertainty in traffic, because the historical state of vehicles are not considered. This paper proposes a novel driving behavior decision-making method EnMFO-ImGRU based on Gated Recurrent Unit (GRU) and Moth-Flame Optimization algorithm (MFO). Four improvements are proposed in EnMFO-ImGRU. First, to consider the driving information of the vehicles on the road, ImGRU is designed based on a double-layer GRU. Second, to promote decisions accuracy, Support Vector Machine (SVM), which has good performance in classification problems, replaces the softmax classifier to train the output of the ImGRU. Third, to promote the classification capability of SVM, MFO is introduced to optimize the key parameters that affect the performance of SVM. Finally, to promote the optimization capability of MFO, we propose the Enhanced Moth-Flame Optimization algorithm (EnMFO). A new position updating method is proposed in EnMFO. The experimental results on the NGSIM dataset show that EnMFO-ImGRU brings higher accuracy than existing methods for the behavior decision-making results of autonomous vehicles.

An Integrated Lateral and Longitudinal Decision-Making Model for Autonomous Driving Based on Deep Reinforcement Learning

“…Since decision-making is an important part that links perception and trajectory planning and greatly determines the safety and efciency of autonomous driving, extensive research around this issue can be found in the literature. In general, the typical research method of autonomous driving decision-making can be mainly categorized into 4 classes: rule-based [3][4][5], classical machine learning-based [6][7][8], deep reinforcement learning-based [9][10][11][12], and deep imitation learning-based [13][14][15]. Among many research methods, deep reinforcement learning has received great attention in recent years because it does not need a lot of human labeled training data, the learning style is closer to human learning, and the generalization ability is strong.…”

CitySim: A Drone-Based Vehicle Trajectory Dataset for Safety-Oriented Research and Digital Twins

“…Currently, a significant portion of related studies have been formulated using video-based trajectory datasets. These studies include driver behavior analysis ( 1 – 3 ), autonomous vehicle virtual testing ( 4 – 6 ), crash mitigation and avoidance system design ( 7 – 10 ), advanced autonomous control algorithm development ( 11 – 14 ), surrogate safety measures ( 15 – 18 ), among other safety applications. In the autonomous vehicle safety field particularly, datasets can provide training/testing samples for personalized algorithms of trajectory/intention prediction and vehicle control, given that the dataset contains rich information about driving preferences.…”