Explicit MPC in the form of Sparse Neural Networks

This paper discusses the construction of a neural network that approximates the behavior of a model predictive control strategy. The aim is to train a neural network that has very similar closed-loop behavior compared to the model predictive control. Furthermore, the paper presents methods to decrease the memory footprint of the neural network-based controller. To achieve the goal, we utilize a genetic algorithm in the training phase, that not only searches for the right weights of the individual neurons, but also for the structural properties of the network. Moreover, the algorithm determines which type of activation function should be considered in which neurons and which neurons should be connected together. The efficacy of the proposed control method is demonstrated on a laboratory scale device Flexy. Moreover, we will show, that presented approaches drastically reduce memory footprint even for simple control problems.