Marco Prattico

Model-free Reinforcement Learning (RL) algorithms have shown significant potential for solving single-task sequential decision-making problems with high-dimensional states and extended time horizons. However, they often struggle to generalize across multiple tasks. In contrast, model-based RL learns task-independent representations of the environment, enabling transfer across different reward functions, though it faces challenges in scaling to complex environments. Recent techniques in modeling dynamical systems, such as Koopman operator learning or Conditional Mean Embeddings offer a promising approach to learn the world model in RL. This project will build on these tools to tackle the generalization limitations of RL by demonstrating the theoretical and empirical effectiveness of these approaches in various application domains, including robotics and neuroscience.