Super Mario Bros.

A reinforcement learning project designed to learn and complete the original Super Mario Bros. for the Nintendo Entertainment System using a Deep Q-Learning model and Asynchronous Advantage Actor-Critic.

Setting up the repository

Creating a virtual environment

After cloning the repository, it is highly recommended to install a virtual environment (such as virtualenv) or Anaconda to isolate the dependencies of this project with other system dependencies.

To install virtualenv, simply run

pip install virtualenv

Once installed, a new virtual environment can be created by running

virtualenv env

This will create a virtual environment in the env directory in the current working directory. To change the location and/or name of the environment directory, change env to the desired path in the command above.

To enter the virtual environment, run

source env/bin/activate

You should see (env) at the beginning of the terminal prompt, indicating the environment is active. Again, replace env with your desired directory name.

To get out of the environment, simply run

deactivate

Installing Dependencies

While the virtual environment is active, install the required dependencies by running

pip -r requirements.txt

This will install all of the dependencies at specific versions to ensure they are compatible with one another.

Training a model

To train a model, use the train.py script and specify any parameters that need to be changed, such as the environment or epsilon decay factors. A list of the default values for every parameters can be found by running

python train.py --help

If you desire to run with the default settings, execute the script directly with

python train.py

The script will train the default environment over a set number of episodes and display the training progress after the conclusion of every episode. The updates indicate the episode number, the reward for the current episode, the best reward the model has achieved so far, a rolling average of the previous 100 episode rewards, and the current value for epsilon.

Any time the model reaches a new best rolling average or a new high score, the current model weights are saved as a .dat file with the environment's name (such as SuperMarioBros-1-1-v0.dat). This saved model will overwrite any existing model weight files for the same environment.

Once the new model is saved, the model will be tested against the requested level to determine the overall performance. The test run is saved in the recording directory and contains a MP4 file of the run to analyze the current performance of the model.

Note: Currently, the testing module throws errors stating the environment is already closed. Despite the errors being thrown, all of the functionality will continue as expected and these can be safely ignored.

Action spaces

This repository allows users to specify a custom set of actions that Mario can use with various degrees of complexity. Choosing a simpler action space makes it quicker and easier for Mario to learn, but prevents him from trying more complex movements which can include entering pipes and making advanced jumps which might be required to solve some levels. If Mario appears to struggle with a particular level, try simplifying the action space to see if he makes further progress.

Currently, the following options are supported:

Right only

Mario can effectively only go right. This simplifies the training process, but prevents Mario from trying more complex actions. The following buttons are supported:

• Nothing
• Right
• Right + A
• Right + B
• Right + A + B

Simple movement

In addition to moving right and running/jumping, Mario can now walk left and jump in place. The following buttons are supported:

• Nothing
• Right
• Right + A
• Right + B
• Right + A + B
• A
• Left

Complex movement

This action allows Mario to try nearly any of his possible actions from the game. This option should be chosen by default for the most realistic exploration of a level, but can increase the time and complexity of learning a level. This is the only provided action space that allows Mario to enter vertically-oriented pipes. The following buttons are supported:

• Nothing
• Right
• Right + A
• Right + B
• Right + A + B
• A
• Left
• Left + A
• Left + B
• Left + A + B
• Down
• Up

Progress

The following table shows the current progress of the model on various levels and the settings used to achieve the indicated performance:

Level	Version	Status	Actions	GIF
World 1-1	v0	Optimal	Complex
World 1-2	v0	Optimal	Simple
World 1-3	v0	Optimal	Simple
World 1-4	v0	Optimal	Simple
World 2-1	N/A	Untested	N/A	N/A
World 2-2	N/A	Untested	N/A	N/A
World 2-3	v0	Optimal	Simple
World 2-4	v0	Optimal	Simple
World 3-1	N/A	Untested	N/A	N/A
World 3-2	v0	Optimal	Simple
World 3-3	N/A	Untested	N/A	N/A
World 3-4	v0	Optimal	Simple
World 4-1	N/A	Untested	N/A	N/A
World 4-2	N/A	Untested	N/A	N/A
World 4-3	v0	Optimal	Simple
World 4-4	N/A	Untested	N/A	N/A
World 5-1	N/A	Untested	N/A	N/A
World 5-2	N/A	Untested	N/A	N/A
World 5-3	N/A	Untested	N/A	N/A
World 5-4	v0	Optimal	Simple
World 6-1	v0	Optimal	Simple
World 6-2	N/A	Untested	N/A	N/A
World 6-3	N/A	Untested	N/A	N/A
World 6-4	v0	Optimal	Simple
World 7-1	N/A	Untested	N/A	N/A
World 7-2	N/A	Untested	N/A	N/A
World 7-3	v0	Optimal	Simple

Legend

The following is a legend of values to decipher the table above.

Level

The level as displayed in the actual game. World 1-1 referes to World 1, level 1 of Super Mario Bros. (ie. the first level).

Version

The version of the environment that was tested. See the Environments section of gym-super-mario-bros' README for examples of the various environment versions.

Status

The current status of training for the indicated level. The status can take on the following values:

• Untested: No attempts or progress has been made on training for the given level yet.
• Training: Training has begun for the indicated level, but Mario has not yet completed the level. If a model is provided, it will correspond to the most recent training pass achieved, and not necessarily the best run so far.
• Satisfactory: Mario can successfully complete the level, but is currently unable to do so in an optimal manner for any reason, including standing in place, losing health, not making forward progress, or others.
• Optimal: Mario has trained enough that he can beat the level at near-optimal performance. This does not necessarily mean the run is perfect, but he can complete the level with only a couple minor interruptions at most. In this state, further progress will likely not be made.

Actions

The action-space that Mario has been trained to use. See "Action spaces" above for more details on the various action spaces.

GIF

An animated GIF of the run that corresponds to the saved model provided in the repository.