Trustworthy Decision-Making Model for Autonomous Vehicles using Graph Neural Networks and Proximal Policy Optimization

In recent decades, autonomous vehicles mostly use advanced systems and Machine Learning (ML) algorithms to make accurate decisions, without involvement of humans. Even though many researchers have developed robust model to enhance the robustness of the system with safety and reliability of autonomous vehicles still there are challenges in handling the events which are unpredictable. Therefore, this research proposes a model namely Graph Neural Networks and Proximal Policy Optimization (GNN-PPO) for decision making of autonomous vehicles. Firstly, the data is collected from Performance Measurement System - Bay Area (PEMS-BAY) dataset. Then, the data collected is preprocessed using Kalman Filter (KF) and min-max scaling for handling noise and normalization. Further, the preprocessed data is embedded into Mutual Information (MI) for optimal selection of features. Finally, the proposed GNN-PPO is incorporated to train with spatial dependencies between traffic sensors and adaptive decision making respectively. The proposed GNN-PPO acquired greater results with low level density return of (198.76 ± 1.06) and high-level density return of (165.30 ± 19.35) respectively when compared to the existing Observation Adversarial Reinforcement Learning (OARL) model.