Quantum-Gated Recurrent World Models: Investigating the Effects of Integrating Quantum RNNs into Model-Based Reinforcement Learning

Model-based reinforcement learning (MBRL) aims to reduce the number of environment interactions by learning an internal model of the environment that supports policy training by providing learned transition dynamics. The fidelity and stability of this model are of crucial importance. While state-of-the-art methods typically employ large, parameter-heavy architectures to capture complex transition dynamics, creating compact yet expressive models remains an open challenge. Recent advances in quantum computing have motivated the exploration of quantum machine learning models as parameter-efficient architectures capable of modeling complex relationships. In this thesis, we investigate the feasibility of employing hybrid quantum-classical architectures to model the recurrent transition dynamics in MBRL. We propose an evaluation framework based on World Models, in which we can seamlessly exchange different variants of recurrent cells in the state-transition model. This allows us to evaluate quantumenhanced recurrent models against a classical baseline with regard to single-step prediction accuracy, multi-step rollout stability, downstream policy performance, and generalization to unseen environment scenarios. All experiments are conducted in a real-world-inspired autonomous driving simulation environment with multimodal sensor input. We use a multi-fold evaluation process and conduct a large-scale hyperparameter optimization. Our results demonstrate that quantum-enhanced recurrent models achieve comparable performance to classical architectures while requiring substantially fewer parameters. Although no clear quantum advantage is observed, our findings demonstrate the feasibility of hybrid quantum architectures for learning latent dynamics and provide a foundation for future research on integrating quantum machine learning into MBRL.