Speaker
Prof.
Daniel Hennes
(University of Stuttgart)
Description
Deep learning techniques allow us to scale reinforcement learning to
problems that were previously intractable, i.e. to domains with
high-dimensional state (or observation) spaces and continuous action
spaces. We will give an overview of state-of-the-art deep
reinforcement learning methods, including deep Q-learning, deep
deterministic policy gradients, and asynchronous advantage
actor-critic. We furthermore show the use of deep neural networks to
approximate the optimal state-feedback control of continuous time,
deterministic, non-linear systems - ranging from toy problems to
fuel-optimal spacecraft landing. The deep neural control policies are
obtained through imitation learning from optimal trajectories. The
method is able to capture the optimal state-feedback with high
accuracy and to generalize well beyond training data.
Primary author
Prof.
Daniel Hennes
(University of Stuttgart)
