HybridNet: Integrating Modelbased and Datadriven Learning to Predict Evolution of Dynamical Systems
[edit]
Proceedings of The 2nd Conference on Robot Learning, PMLR 87:551560, 2018.
Abstract
The robotic systems continuously interact with complex dynamical systems in the physical world. Reliable predictions of spatiotemporal evolution of these dynamical systems, with limited knowledge of system dynamics, are crucial for autonomous operation. In this paper, we present HybridNet, a framework that integrates datadriven deep learning and modeldriven computation to reliably predict spatiotemporal evolution of a dynamical systems even with inexact knowledge of their parameters. A datadriven deep neural network (DNN) with Convolutional LSTM (ConvLSTM) as the backbone is employed to predict the timevarying evolution of the external forces/perturbations. On the other hand, the modeldriven computation is performed using Cellular Neural Network (CeNN), a neuroinspired algorithm to model dynamical systems defined by coupled partial differential equations (PDEs). CeNN converts the intricate numerical computation into a series of convolution operations, enabling a trainable PDE solver. With a feedback control loop, HybridNet can learn the physical parameters governing the system’s dynamics in realtime, and accordingly adapt the computation models to enhance prediction accuracy for timeevolving dynamical systems. The experimental results on two dynamical systems, namely, heat convectiondiffusion system, and fluid dynamical system, demonstrate that the HybridNet produces higher accuracy than the stateoftheart deep learning based approach.
Related Material


