The field of Inverse Reinforcement Learning (IRL) addresses the task of finding a cost function which describes expert behavior. Since the cost function is solely computed from expert demonstrations, the sample complexity exerts influence on the performance of these algorithms. In this thesis we study the Learning to Search (LEARCH) and the maximum entropy IRL framework as example IRL techniques. Based on these two algorithms we develop a variation of the LEARCH algorithm using the idea of maximum entropy IRL. In the next step we extend LEARCH to Deep-LEARCH as well as the newly developed LEARCH variation to a equivalent Deep-LEARCH variation. Thereby we generalize the cost function to function space using Convolutional Neural Networks (CNNs). Including maximum entropy inside the Deep-LEARCH variation increases the density of the target maps of the CNN. We discover that LEARCH shows the lowest sample complexity among the investigated algorithms, while maximum entropy shows the highest sample complexity. In the deep learning setting the increased density of the CNN target maps did not improve the performance. Hence, the performance does not change, if the algorithms are extended by CNNs to the function space.