Explainable artificial intelligence (xai) for deep reinforcement learning based autonomous driving

In recent years, there has been a notable surge in the development of autonomous vehicle technology, driven by the rapid advancements in artificial intelligence and computer technologies. At the current time, several key industry stakeholders are engaged in active work on the development of autonomous vehicles. While the accelerated advancement of autonomous driving technology through data-driven and deep learning-based techniques has the potential to mitigate traffic issues and enhance efficiency, the intricacy and black box structure of these models present a considerable obstacle to the comprehensibility and transparency of decision-making processes. In this context, a number of legal, ethical and socio-psychological issues emerge. The application of explainable artificial intelligence (XAI) methodologies has emerged as a promising field of study with the potential to enhance the reliability, accountability and transparency of autonomous driving systems. In this context, XAI has the potential to play an important role in ensuring compliance with legal regulations, increasing user confidence and identifying deficiencies in software development. This thesis analyses and evaluates the performance of deep reinforcement learning (DRL) based autonomous driving systems and XAI approaches under different scenarios. Furthermore, the necessity for XAI methods to guarantee that autonomous vehicles comply with legal regulations and are accepted by society is emphasised. The TD3, DDPG and PPO methods were selected as the DRL models. The experiments conducted in the TORCS simulation environment are presented in a comprehensive manner, including the training and performance evaluation of DRL models and SHAP analyses. The findings obtained from the training and testing phases indicated that TD3 was the most successful model. The impact of specific features on the decision-making processes of the models was evaluated through global and local explainability analyses utilising SHAP techniques. In the local explainability analyses of the TD3 model, it was observed that the TD3 model exhibited actions that were comparable to those of expert human drivers. The obtained findings contribute to the advancement of the reliability and accountability of autonomous vehicles. This thesis demonstrates that XAI approaches play a pivotal role in ensuring that autonomous driving systems comply with both technical and legal requirements.