2019/11/04 by S. Pozzoli, Pozzoli, Simone, Marco Gallieri +3
Computer Science · Engineering · #Advanced Control Systems Optimization #FOS: Computer and information sciences #FOS: Electrical engineering #Fuzzy Logic and Control Systems #Neural Networks and Applications #Neural and Evolutionary Computing (cs.NE) #Systems and Control (eess.SY) #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1911.01310
openalex publication_date 2019/11/04 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
The use of recurrent neural networks to represent the dynamics of unstable\nsystems is difficult due to the need to properly initialize their internal\nstates, which in most of the cases do not have any physical meaning, consequent\nto the non-smoothness of the optimization problem. For this reason, in this\npaper focus is placed on mechanical systems characterized by a number of\ndegrees of freedom, each one represented by two states, namely position and\nvelocity. For these systems, a new recurrent neural network is proposed:\nTustin-Net. Inspired by second-order dynamics, the network hidden states can be\nstraightforwardly estimated, as their differential relationships with the\nmeasured states are hardcoded in the forward pass. The proposed structure is\nused to model a double inverted pendulum and for model-based Reinforcement\nLearning, where an adaptive Model Predictive Controller scheme using the\nUnscented Kalman Filter is proposed to deal with parameter changes in the\nsystem.\n