Given the successes of convolutional neural networks in visual processing, the next step is to use the power of deep learning to enhance the abilities of agents
which affect their environment through their actions. To help the community, we have worked on software
to enable state-of-the-art deep reinforcement learning algorithms to be applied to different domains. This work has been featured on the Torch blog
as part of an introduction to reinforcement learning and the deep Q-network.
From a theoretical perspective we are interested in the benefits of different neural network architectures [Arulkumaran et al, 2016], with respect to both hierarchical reinforcement learning and transfer learning; this work is done closely with the Computational Neurodynamics Group in the Department of Computing. We are also investigating practical applications in end-to-end learning of visuomotor control policies for robots.