[PYTORCH] Reinforcement Learning-Super Mario Bros Using CNNPolicy

Packages Used:

• gym-super-mario-Bros == 7.3.2
• gym == 0.21.0
• ipykernal
• Stable-Baselines3
• Pytorch 1.13

Setting Up Pytorch

I would recoomend you to check official Pytorch website to install it
The Pytorch need the exact version of Cuda and CUDnn to work normally
Pytorch website: pytorch

Creating Virtual Environment:

setting up a virtual environment will really help in Creating a seperatiion wwithin your workspace and PC
Run following to create a virtual environment

• launch cmd
• move to your desired directory
• Run the following: python -m venv mario activate your virtual env using following:
  .\mario\Scripts\activate
  to deactivte just type deactivate in cmd

Creating a New kernel in Jupyter Notebook

After installing Ipykernal run the following:
python -m ipykernel install --user --name=mario

note: the name of kernel should be same as venv name

Training the Agent

Either clone my repo or download the zip or copy paste the code and run them
Github clone Link: https://github.com/harikris001/Super-Mario-Reinforcement_Learning.git
Run the cells and wait for it to complete.

After Training:

OUTPUT
Sl.NO	1	2	3
Total Steps	.5M	1M	1.5M
Completed Level	False	False	False
OUTPUT
-----------------	------	------	------
Sl.NO	4	5	6
Total Steps	2M	2.5M	3M
Completed Level	False	False	False
OUTPUT
-----------------	------	------	------
Sl.NO	7	8	9
Total Steps	3.5M	4M	4.5M
Completed Level	False	False	False
OUTPUT
-----------------	------	------
Sl.NO	10	11
Total Steps	5M	5.5M
Completed Level	False	False

I have attached the Pre trained models u can resume training using these model and save the result
The best model during my training was 4M model so i have attached that