392284 ISY Project: Unsupervised pretraining & transfer learning in pixel based deep reinforcement learning (Pj) (SoSe 2018)

Pixel-based end-to-end reinforcement learning has recently attracted a lot of attention, particularly
as a result of the success in gaming (published by a Google team for Atari games in Nature) and
subsequent efforts follow in these footsteps. A drawback of such end-to-end learning for image-based
signals is that the internal representation of data is learned afresh for every new game. The goal of
the project is to investigate whether it is possible to learn a more universal representation by means of
deep learning (such as deep autoencoders) which enables to leverage end-to-end learning for
specific games by its particularly efficient initialization. Such transfer could leverage the efficiency
of reinforcement learning for one specific game but possibly also along different games where different
visual features ight be of importance. Within the project, the task is, based on a basic end-to-end RL
chain and standard autoencoders, to elucidate suitable scenarios in which such transfer of the learned
representation can be tested.

- In case this would not find enough interest for a team project, this project proposal would be also offered (in reduced/modified form)

The project team for this project has already been selected, no additional applications are considered.

Requirements for participation, required level

- Required skils (e.g. mandatory courses, if required)
Advanced Machine Learning skills and knowledge about deep learning technology (eg tensor flow) are required.