HBP Curriculum: Interdisciplinary Brain Science | Cognitive systems for non-specialists | 4th Teaching Cycle
Lecture 5: Deep reinforcement learning for robotic control
Speaker: Jonathan Hunt, Google DeepMind, UK
Combining the representation capacity of deep neural networks with reinforcement learning (Deep RL) has lead to a number of notable successes. RL in continuous action spaces brings its own unique set of challenges, as it is no longer possible to enumerate all possible actions. I will discuss several recent scalable approaches to RL in continuous action spaces including Deep Deterministic Policy Gradient, an off-policy method we developed. I will also discuss limitations of these approaches, particular relating to the lack of transfer learning between tasks. The second half of the talk I will discuss the successor representation. Successor representations decompose the value function into a reward independent part and a reward dependent part, which allows it to generalize across tasks and, despite learning with model-free methods, regain some of the benefits of model-based approaches. I will discuss new, scalable innovations using successor features.
Visit our HBP Curriculum: Interdisciplinary Brain Science website: https://www.humanbrainproject.eu/en/education/participatecollaborate/curriculum/
Discuss with us, the lecturer or other participants in our forum: https://forum.humanbrainproject.eu/