This is a project I did in the context of my Master's in Artificial Intelligence at the Unviersity of Amsterdam.
The main report can be found in the root repository directory.
Mainly I am investigating the effect of different representational models, such as VAEs, VQ-VAEs/dVAEs and standard CNNs, as well as implementing and testing an MDN-RNN (based on World Models (Ha & Schmidhuber, 2018). The final RL agents are trained via DQfD, which is also reimplemented.