Vast amounts of clinical data, which often does not exist, are required for developing deep learning models. In this work, we focus on the problem of synthetic data generation in the context of hospital infection spread and contacts. We identify 3 key challenges: (1) epidemiological compartmental models are the gold standard in medical literature, focusing on simulating populations but ignoring individual patients and hospital conditions. (2) Agent-based models can simulate patient’s behavior and their interactions with the hospital setting using stochastic rules. However, they rely on several hyperparameters that require manual tuning. (3) Both approaches are rarely merged, but, in such cases, they are highly sensitive on a wide range of hyperparameters. There is a growing interest in the use of reinforcement learning (RL) for building non-parametric models since it mainly needs a tailored reward function. However, limited research has been done for clinical synthetic data and spatial-temporal information. We explore the potential of RL to optimize a simulation model for Clostridioides difficile spread in hospitals. In our experiments, the model is learnt from a synthetic dataset validated using an agent-based system (2); RL is driven by a reward function based on the adherence to the SEIRD epidemiological model (1); the final model obtained is non-parametric and human interpretable (3). We have experimented with 5,049,323 experiences to achieve convergence of the model and adherence to the SEIRD. The experiments show promising results in the application of RL to bring a stochastic model closer to the behavior of a compartmental model.