World Models - Can agents learn inside of their own dreams?#
Humans build an internal mental model of the world and routinely use simulation to update it state. The work in World Models uses RNNs to tackle Reinforcement Learning (RL) tasks and divide the agent into a substantial world model and a smaller controller model.
Task 1 (20 points)#
Read the paper and write your own 4-page report of the technique (a model-based RL), in a tutorial like fashion so computer scientists can still understand it.
Task 2 (40 points)#
In this task you are asked to reproduce the results for the car racing environment. Consider adopting this repo (30 points).
Add to the report of a section that documents your own results and comment on the speed of learning a policy that drive the car around the track (10 points)
Task 3 (40 points)#
Familiarize yourself with Generative Adversarial Networks (GANs). Example destinations:
Document a new approach that combines the VAE with a GAN (20 points)
Repeat the experiment in the car racing environment. There are multiple implementations of the VAE/GAN approach. Add to the report what improvements (if any) you observed from the transition to GAN in terms of performance (20 points)