Developing Improved Perception for Reinforcement Learning Agents in Complex Environments
In this project, the plan was to develop a method of perception for reinforcement learning agents which may allow them to operate within seemingly complex environments. In a recent paper, one such architecture of perception was proposed called the World Models. This method constitutes of three components namely, Vision(V), Memory(M), and Controller(C) models. The V model uses a variational autoencoder made from of CNNs to create a compressed representation of the environment. The M model uses mixture density networks made from RNNs in order to give the whole architecture the ability to learn from its previous states. The C model uses evolutionary strategies to control the agents actions according to the repre- sentation it receives from the previous two models. This whole architecture has already been implemented on two relatively simpler gaming environments, CarRacing and DoomTakeCover, both by the creators of the World Models. Implementing this in a more complex environment and observing the results obtained is the objective of this project.