This study presents an investigation into the application of Imitation Learning (IL) techniques for extracting structured and interpretable policies from black-box expert models. The primary focus is to analyze the feasibility and effectiveness of this approach in translating behaviors learned by deep neural networks into decision trees, which represent a set of rules that can be sequentially evaluated to reach a decision. We evaluate this methodology in three distinct simulation environments: Lunar-Lander, Taxi, and CartPole. We test the DAgger algorithm and its variant VIPER, which iteratively train policies represented by decision trees from demonstrations of an expert policy. We compare the use of traditional decision trees with linear model trees, which contain linear models in their leaves.