Data Efficient, Robust, and Interpretable Deep Reinforcement Learning for Robotics and Dynamical Systems

Deep reinforcement learning has achieved outstanding success in learning optimal control strategies directly from high-dimensional data by exploiting the expressive power of neural networks. However, neural network-based control policies tend to be over-parametrized, which means they need large amounts of training data, show limited robustness, and lack interpretability. These drawbacks become even more severe when controlling dynamical systems requiring real-world interaction and expensive simulation platforms, e.g., robots and fluids. In this talk, we will look at how to improve sample efficiency, robustness, and generalization in deep reinforcement learning through (i) unsupervised learning of low-dimensional data representations for robotics, (ii) sparse dictionary learning of control policies for parametric partial differential equations, and (iii) neural ordinary differential equations to blend physics and machine learning and to achieve energy-efficient control.