AutoEncoder for Latent Representation Learning and Multi-Agent Coordination using DDPG

• First we use an autoencoder to compress the state representation to one-fourth the original size.
• Then we train DDPG agents to learn joint optimal policies to observe points of interests in a rover domain using the compressed state representation.
  • The learned policies perform as good as the policies learned with the original state representation.
    

P.S. This work uses a codebase of the rover environment from Autonomous Agents and Distributed Intelligence (AADI) Lab from Oregon State University. Hence, this repo includes some code not used in this project.

Demo