SLWM: A Library for Implementing Complex Training Workflows for surrogates of MPC’s

Model Predictive Control (MPC) is an advanced control technique that uses a model to predict the system’s future states to optimize its control actions. However, it can be computationally intensive and challenging to provide real-time guarantees. Controller surrogates, based on machine learning techniques, try to overcome this barrier. Training neural networks as surrogates for MPCs often involves intricate workflows, including data generation, model training, and iterative refinement, which can be challenging to manage efficiently. In this paper, we introduce SLWM, a Library that helps create complex workflows used when training a (deep) neural network to act as a surrogate of an MPC. Through a case study of an inverted pendulum, we demonstrate its effectiveness in implementing workflows like Dagger. The results of this implementation show that SLWM fulfils the requirements such as modularity, readability, simplified data management and reliability. Lastly, we list a few potential future applications of SLWM and the extensions of SLWM required for some of them.