Brief description of the PhD project:

PhD on the topic of "Explainability, Control, and Learning Efficiency in Agentic AI Systems" [working title]

My doctoral research focuses on advancing the foundations of Agentic AI systems, with a particular emphasis on explainability, controllability, and learning efficiency. Unlike classical decision‑support systems, which provide information or recommendations for humans to act upon, Agentic AI systems make autonomous decisions and directly interact with their environment. This shift from supporting decisions to taking decisions introduces new requirements: these systems must not only be transparent, but also predictable in their behavior, aligned with human or organizational objectives, and robust under uncertainty, as their actions have immediate real‑world consequences. Agentic AI systems consist of autonomous decision‑making entities—AI agents—whose behavior is typically learned through Deep Reinforcement Learning (DRL). As these systems are increasingly deployed in complex, uncertain environments, understanding and influencing their internal reasoning processes has become a central scientific challenge. The ability to interpret how an agent forms internal representations, how it evaluates uncertainty, and how it adapts to dynamic conditions is essential for ensuring safe and reliable autonomous behavior. A core objective of this research is to develop methods that make the learning dynamics of Agentic AI systems more transparent and more aligned with human or system‑level goals. This includes studying how noise, stochasticity, and instability in the environment propagate through the learning process, and how these factors affect policy formation. By integrating explainability with control‑oriented mechanisms, the project aims to create systems whose behavior can be monitored, guided, or constrained without compromising performance. Another major focus lies in addressing the inefficiencies and noise inherent in DRL training. Traditional DRL algorithms often suffer from high variance, unstable gradients, and slow convergence, particularly in high‑dimensional or partially observable environments. The research explores strategies to mitigate these limitations through noise‑reduction techniques, and architectural modifications that enhance sample efficiency. The overarching goal is to design Agentic AI systems that learn significantly faster while maintaining robustness against uncertainty and dynamic changes. Overall, this doctoral project aims to deepen the scientific understanding of how Agentic AI systems learn, reason, and behave—and to develop methods that make these systems more transparent, predictable, controllable, and efficient. The long‑term vision is to contribute to the creation of AI systems that are not only powerful but also trustworthy and aligned with human decision‑making needs.