Monday, December 18, 2023, 15:00
Room 01-012 in Georges-Köhler-Allee 102
As deep reinforcement learning techniques have recently become more powerful, and produced significant achievements in decision making problems, their use has largely been limited to simulated environments. Leveraging their power to learn from data in the real-world offers significant benefits. However the unsafe nature of data-driven methods remains a hazard in the real-world where applications are safety-critical, and involve interaction with the humans. Finding approaches to make these methods safer are of critical importance before they can be widely employed in the real world. Using expert demonstrations can make the learning process safer in applications where learning by trial-and-error is not affordable. In this thesis, we apply behavioral cloning, which is an imitation approach, to learn to drive a car around a race track using a neural network, by learning from expert demonstrations of a model predictive controller. Several formulations of neural networks are tested for their effectiveness in learning from the demonstration trajectories. Finally, the performance of the neural network is compared with the model predictive controller in terms of computation cost, lap time, and reliability. The neural network not only reduces the computation time by half, but also provides
marginally faster lap times when used with a safety filter.