ZhangXi-2020/EI339-RL-Project

Team project about reinforcement learning for SJTU EI339-Artificial Intelligence

EI339-RL-Project

Team project about reinforcement learning for SJTU EI339-Artificial Intelligence.

In this project, we implement two tasks about reinforcement learning.

• Easy 21 game
• Reinforcement learning in Quanser Robot platform

Easy 21 game

We apply Q-learning algorithm and policy iteration method to solve Easy21 problem.
Some learning result:

Reinforcement learning in Quanser Robot platform

We implement the TRPO and MPC methods on Qube, Ball Balancer, CartPoleSwing platforms and solve the problem of each environment.
For the installation of the Quanser robot simulation environment, please see https://git.ias.informatik.tu-darmstadt.de/quanser/clients.

Some hyperparameter experiments figures:

Reference:

• TRPO paper: https://arxiv.org/abs/1502.05477
• TRPO code: https://github.com/ikostrikov/pytorch-trpo
• MPC paper: https://ieeexplore.ieee.org/abstract/document/8463189
• MPC code: https://github.com/michaelliyunhao/RL-project

How to run

1. Easy 21 game
   • Q-learning algorithm

   python .\1-Easy21\Q-learning\run_q_learning.py
   

   • policy iteration method

   python .\1-Easy21\Policy-iteration\run_policy_iteration.py
   

2. Quanser Robot platform
   • TRPO

   python .\2-Quanser_Robot\TRPO\main.py --env-name "Qube-100-v0"
   

   • MPC

   python 2-Quanser_Robot\MPC\MPC-Qube\run.py