Two-Layer-Reinforcement-Learning

Two Layer Structure : Reinforcement Mechanism Design for Multi Agent System

Outer Layer - Mechanism Designer :

• Implemented based on open-ai gym environment
• Mechanism Designer chooses a mechanism parameter as action
• Step Function is a full execution of the multi agent system (aka, inner layer)
• Mechanism Designer evaluates its actions based on a predefined reward function

Training Outer Reinforcement Learning Problem

• Single Agent Reinforcement Learning Problem
• Recommendation : Train using TD3 Algorithm from Stable Baseline Library

Inner Layer - Multi Agent System :

• Implemented based on petting-zoo MPE
• Simple Multi Agent Scenario
  • each agent collect food & gather score during each episode
  • can become more complex by adding other elements to the game (barrier, bomb, ... )

Training Inner Reinforcement Learning Problem

• Multi Agent Reinforcement Learning Problem
• Two approaches for training :
  • MADDPG : Multi Agent DDPG
  • DDDPG : Distributed DDPG (or other single agent environments)
  • for comparison of these two approaches check maddpg paper