Explainable end-to-end automated driving system using counterfactual explanations

Harnessing the performance of Deep Neural Networks (DNNs) in safetycritical applications, such as Automated Driving Systems (ADSs), requires a clear understanding of their reasoning processes to explain why specific decisions are made under given circumstances. This thesis focuses on developing explainable deep learning techniques, with an emphasis on Counterfactual (CF) explanations for ADS. CF explanations identify minimal changes to input data that would alter a model’s output, providing insights into the rationale behind decisions by highlighting the influence of specific input elements. The research begins by introducing SAFE (Saliency-Aware CF Explanations), a novel framework that integrates saliency maps with a Generative Adversarial Network (GAN) to generate informative CF explanations for supervised DNNs. SAFE advances current methods by emphasising discriminative features and producing CFs that lie near decision boundaries. Its effectiveness is validated through extensive experiments on the BDD100k driving dataset. Building on this foundation, the second study extends SAFE to Deep Reinforcement Learning (DRL) environments, particularly in Automated Driving (AD) and gaming. By incorporating innovative modifications to the GAN architecture, the resulting framework, SAFE-RL, ensures that generated CF explanations are both plausible and temporally consistent across sequential observations. This extension yields sparse and practical CF examples, enhancing the interpretability of DRL model decisions. The third study introduces the Policy-Distillation CF Explainer (PDCFE), a framework specifically designed for explaining DRL model decisions. PDCFE addresses the challenge of generating CF explanations for high-dimensional inputs by employing policy distillation to map DRL policies into a low-dimensional space. The framework generates CF explanations using gradient-based methods before translating them back to the high-dimensional space. Experimental results in both AD and gaming environments validate PDCFE’s ability to generate meaningful and plausible explanations. The final study presents the CF-Driver framework, which utilises CF examples as a data augmentation technique for imitation learning in ADS. By incorporating CF scenarios, CF-Driver improves handling of rare situations and boosts CARLA driving scores by up to 15.02% over state-of-the-art models. By bridging the gap between complex AI decision-making and humanunderstandable reasoning, this research advances explainable AI for ADS. The findings contribute to safer and more interpretable autonomous systems, supporting their real-world deployment and fostering public trust.