Master Thesis Approximating Model Predictive Controllers Using Imitation Learning

Approximate model predictive control (AMPC) has emerged as an approach to tackle the computational burden of MPC, aiming to approximate the MPC policy with a computationally cheaper surrogate, such as, e.g. neural networks. So far, the standard approach to obtain such a surrogate policy is based on naive behavioral cloning. This approach, however, has significant drawbacks, resulting in the surrogate policy to potentially not provide the original MPC guarantees. To tackle this, a tailored AMPC imitation learning (IL) procedure was developed recently, enabling consistent learning of a surrogate policy, and ensuring that the learned policy maintains the original MPC safety and stability guarantees, thereby enabling MPC-based control functions in safety critical industrial settings.

• The goal of your thesis is to extend the statistical properties of the proposed IL procedure by analyzing the rate with which the learned policy converges to the MPC policy, ultimately with the goal of providing finite sample bounds on the error between the policies.
• Moreover, the thesis could cover the investigation of more generic error estimations, stopping criteria and studies on sample efficiency.
• Based on this, the second goal of your thesis is the deployment of the developed AMPC IL procedure to a real-world automated driving problem. This includes comparison with other existing approaches.

• Education: Master studies in the field of Cybernetics, Engineering, Mathematics, Computer Science or comparable
• Experience and Knowledge: profound knowledge of Machine Learning and Control Engineering; experience with Python DL frameworks such as PyTorch, TensorFlow or JAX
• Personality and Working Practice: you excel at working autonomously, systematically organizing your tasks, and applying analytical thinking to solve complex problems
• Languages: very good in English