Deep Multi-Agent Reinforcement Learning

AY2020/21 Sem 1/2 CP3209 UROP in Computing Project with Dr Jing Wei, IHPC.

To-do list

High Priority

• Get models working for speaker_listener, followed by the rest of the scenarios
• Add discrete action space output option via Gumbel-Softmax reparameterization trick
• Move noise parameter to inside the agent class
• Add support for individual good/bad agent policies
• Implement M3DDPG algorithm
• Implement GIF saving for MPE
• Implement policy estimation and esembling for MADDPG

Medium Priority

• Add support for MultiBoxDiscrete action space
• Add individual agent reward tracking
• Experiment with additional normalization layers
• Experiment with separate actor/critic networks

Low Priority

• Re-factorization of code into packages
• Modify MPE code to provide benchmark statistics
• Document code
• Add ability to set random seed
• Add printout for model and experimental parameters before code execution