Motivated by cam profile generation in mechatronics, we investigate function approximation by Ck-continuous splines using gradient descent optimization provided by the machine learning framework TensorFlow.
First, we investigate the convergence behavior of the model parameters of a spline model with respect to different loss functions, specifically for L2-approximation error and Ck-continuity, by means of different TensorFlow optimizers. In particular, we propose a parameter regularization that allows SGD (with/without Nesterov-Momentum) to converge. We show that Adam (or rather AMSGrad) generally performs best, yet the rather simple SGD can be close to competitive with our regularization.
Wann? MI 2. Feb 2022, 16:00
Wer? Hannes Waclawek, Junior Researcher am Studiengang Informationstechnik & System-Management
Der Vortrag findet im Rahmen der Reading Group der "Smart Factory & Edge Computing"-Forschungsgruppe des Studiengangs Informationstechnik & System-Management statt.