qh1874 / stable-multi-agent-reinforcement-learning

Implementation for the paper "Multiagent Learning Using a Variable Learning Rate"

More Stable Multi-Agent Reinforcement Learning

This project aims to test/implement various algorithms that are related to Multi-agent RL to see whether these algorithm can lead the agent to more stable traing and/or desire behavior (Nash Equilibrium)

Right now we test most of the algorithms on Iterated Prisoner's Dilema to see whether tit-for-tat behavior arises from these kind of training.

Algorithms

Right now, I have planned to implement 3 algorithms

1. Multiagent learning using a variable learning rate (https://www.sciencedirect.com/science/article/pii/S0004370202001212)
2. Consensus Optimization from https://arxiv.org/abs/1705.10461
3. Learning with Opponent-Learning Awareness (https://arxiv.org/abs/1709.04326)