Autonomous Vehicle Decision-Making Using Generalizable and Adaptive Explainable Network

Nowadays, the Autonomous Vehicle (AV) decision-making depends on Artificial Intelligence (AI) models which process sensor data to perceive surroundings, predict outcomes, and make informed decisions. These models allow vehicles to navigate complex environments, recognize objects and respond to dynamic situations. However, the existing eXplainable AI (XAI)-based models failed to generalize in unseen driving scenarios which led to incorrect or unsafe decisions. Hence, this research proposes Generalizable and Adaptive eXplainable Network (GAX-Net) to enable safe and adaptive decision-making in open-world AV environments. Initially, the input data is collected from vehicle-mounted sensors and preprocessed with Extended Kalman Filter (EKF) for sensor fusion. Also, the noise filtering and outlier removal is performed with median filter, statistical outlier removal and random sample consensus. Then, perceiver <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{I O}$</tex> is employed to extract features by unifying data from different sensors into a common representation. After that, Shapley Additive explanation (SHAP) is used to select features by identifying which influences model predictions. Then, Proximal Policy Optimization (PPO) is used to learn optimal driving actions while accounting for uncertainty and safety. Finally, Elastic Weight Consolidation (EWC) is utilized to update the model continuously without forgetting previously acquired knowledge. The proposed GAX-Net achieved better results in terms of accuracy (0.89) than existing XAI-SecureNet framework.