README.md

IDATT2502 - Project Assignment

• written by Daniel Skymoen, Edvard Schøyen, Jarand Romestrand, Kristian Vaula Jensen

Assignment

Various reinforcement learning tasks

Assignment proposals:

• Choose an environment - test different algorithms
• Choose an algorithm - test different environments
• Create an environment
• Explore different configurations of an algorithm
• Try a hardcore environment: MineRL, Atari, Doom​

We chose to try a hardcore environment, and the environment we ended up with was Super Mario Bros.

Read more about the environment here: https://pypi.org/project/gym-super-mario-bros/

Problems statement

Compare the performence of DDQN and PPO in the Super Mario Bros environment.

Requirements

Here are the requirements to run the project:

• Python >=3.9, <3.12

Preferred requirements for easier setup and running of the project:

• Make

Windows

• Visual Studio C++ build tools

Help

To list all the Make commands use the command under:

make help

How to run

The guide for running the project assumes that you have Make installed. If you don't have it, you will find the commands you need in the Makefile

Setup

This function creates a virtual environment for all the dependecies required for running the project and installs them.

make setup

Python3 is set to standard and if your computer uses python instead of python3 you can overwrite this with the command below:

make setup PYTHON=python

Neptun

An .env file is required to use Neptun for logging the training of a model. This must be put inside the src folder. Example of the contents needed for Neptun:

NEPTUNE_API_TOKEN="YOUR_API_KEY"
NEPTUNE_PROJECT_NAME="YOUR_NEPTUN_PROJECT_NAME"

Running the project

The commands under activates the environment created in the setup section and then runs the python code.

Train the models

This function activates the environment and then trains the ddqn.

make ddqn

This function activates the environment and then trains the ppo.

make ppo

Render trained models

This function activates the environment and then renders a trained model of ddqn.

make render-ddqn

This function activates the environment and then renders a trained model of ppo.

make render-ppo

Options

You can specify flags for both the ppo and ddqn to log the training to Neptun. You can use these args on both ppo and ddqn. The args can also be combined. To specify flags for logging use the commands under:

Log training

Logs the training to make graphs on Neptun

make ddqn args="--log"

Log the completed model

Logs the model to Neptun

make ddqn args="--log-model"

Log training and the completed model

Logs the training to make graphs and the model on Neptun

make ddqn args="--log --log-model"

Python options

Pyhton3 is set as standard so if your computer uses python instead of python3 you can overwrite this with the command below:

make ddqn PYTHON=python

Deactivate the environment with:

deactivate

Cleanup

This commands removes the virtual environment created in the setup section. You can also remove this manually.

make clean

Example runs

Here are some example runs:

DDQN

Example runs with DDQN:

PPO

An example run with PPO:

Performance graphs

Here are graphs of the models performance in the environment with 15 000 episodes:

Reward per episode

Stage 1-1

Stage 6-4

Completion rate/flag average the last 100 episodes

Stage 1-1

Stage 6-4