No description, website, or topics provided.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


This repository holds our work for the OpenAI Retro Contest.



  • 64-bit Architecture

How To Do Stuff Relevant To This Project

Table of Contents


There are some novel elements to this project so a decisive vocabulary is necessary for effective communication and cohesive code.

NOTE: These definitions are our own; other documentation often has inconsistent vocabularies. This vocabulary is generally in alignment with the white paper for the contest.

  • Environment: An object instance returned from a call to retro_contest.local.make for local environments, gym_remote.client.RemoteEnv for remote environments, or some wrapper. This object handles the game simulation and provides the interface to it through methods like .reset(), .step(), .render(), and others. A remote environment is created and managed by a Docker container in the background; this is done when submitting an agent to the contest or simulating an agent submission. A local environment is created and managed by the agent itself.
  • Episode: One attempt at completing a level. An episode ends when the player dies, completes the level, or runs out of timesteps (4500). This is signalled in the done value returned by environment.step(). An episode begins with a call to environment.reset() which should only be called immediately after environment creation or when done is True.
  • Agent: A script that simulates episode after episode of some level for some number of timesteps. If the agent uses a local environment then it is responsible for setting and abiding by the timestep limit and outputting a monitor.csv. If the agent uses a remote environment, then it doesn't need to worry about the timestep limit; it can operate as though it's running infinitely and silently while the limit and monitor.csv output is taken care of remotely.
  • State: Imagine the common understanding of level in the context of videogames. Example: Green Hill Zone - Act 1 is the first state of the game Sonic The Hedgehog. The begin far on the left and are considered complete when some horizontal displacement is achieved.
  • Game: A collection of levels and control definitions. These are implemented via ROMs that are copied directly from the original game cartridges. For instance: Sonic The Hedgehog 2 has a Spin Dash move, but Sonic The Hedgehog does not.
  • Trial: A trial is defined by an agent, a level, and a timestep limit. During a trial the agent will simulate play in that level, restarting as necessary, until the timestep limit is reached.
  • Timestep: The smallest unit of time in a simulation. A timestep occurs when some action is passed to environment.step(); the action is enacted for four real game frames (+/- 1) during one timestep. One timestep is approx. 1/15 of a second of wall-clock time in the game, so the episode limit of 4500 timesteps is approx 5 minutes of real gameplay.
  • Test Set: A list of game-state tuples for the purposes of experiments and testing in general.
  • Experiment: An experiment should be designed to explore the behavior of a range of values for a single parameter of an agent. An experiment requires an agent, values for the static parameters of that agent, a range of values for the dynamic variable of that agent, a test set, a trial limit, and a timestep limit. An experiment will run trials according to the timestep limit until the trial limit is exceeded for every value in the dynamic variable range and for every level in the test set (think nested loops).
  • Observation: The graphical data of a frame. This is the first element in the tuple returned by a call to environment.step(). It is a 320x224x3 array of RGB pixel data.
  • Action: An 12 element array of binary values (0/1 or True/False) corresponding to buttons on a Sega Genesis controller such that ON means pressed and vice versa. The mapping of index to button is {0: JUMP, 1: JUMP, 2: (ignored), 3: (ignored), 4: UP, 5: DOWN, 6: LEFT, 7: RIGHT, 8: JUMP, 9: (ignored), 10: (ignored), 11: (ignored)} (Sonic does not make use of all the buttons, and there are redundancies for jumping). An action is performed by passing as an argument to environment.step().
  • Action Space: The set of possible actions to take in an environment.
  • Reward: After every action (a.k.a. call to environment.step(), a reward for the action is presented by the environment as the second item in the tuple return from environment.step(). This magnitude of this reward is proportional to the horizontal displacement achieved by the action and positive if to the right and negative if to the left.
  • Score: Every episode is given a score at the end. This score is the sum total of all the rewards from the actions that took place during the episode (max 9000) plus a bonus for speed (theoretical max 1000). Every trial is given a score which is the mean of the scores of all the episodes that completed. The contest evaluation system awards an agent a total score which is the mean of the episode scores for every level it tests.
  • Container / Image: See this Stack Overflow post. In this documentation "container" is used exclusively and probably not precisely. It's not that important for our purposes.


  • You're keeping everything for the project in a directory called sAnIc. This is the default after having done git clone for this repository or a fork of it.
  • ROMs are stored in a directory called roms under sAnIc, i.e. at sAnIc/roms
  • You're using a shell thats "bash-compatible". That means no csh, tcsh, ksh, etc. The scripts are targeted at POSIX sh, but using bash is the safest choice.
  • You're developing in Python 3. Don't use Python 2. Dont use pip or pip2; use pip3.6 just to be explicitly compliant. Use #!/usr/bin/env python3.6 for your shebangs.
  • You've installed the dependencies yourself. You need how to figure this out for your own environment, e.g. Ubuntu will need calls to $ sudo apt-get install or something. There are some helpful links and hints in the Dependencies section.

Contribute to the Project

Things can get really messy with dependencies and outside libraries and everything else. For that reason we will use pull requests to contribute work instead of straight pushes to the repo.

  1. Fork The Project
    • Visit the main project repo and click Fork just to the left and below the Sanic icon.
    • This may give you a list of entities to fork it with, if so then click the entity representing your personal account.
    • This should create new repo just like the original, but owned by you, and it should have taken you to that repo's main page. If not try to find the new repo in your account's list of repositories.
    • Now clone this repository of yours onto your local machine with the link provided just below the icon.
  2. Add an Upstream Remote
    • Adding this allows you to stay up to date by fetching code from the main project repo and merging it into your forked repo
    • cd to your local copy of the forked repo and run
      $ git remote add upstream
  3. Update the Fork
    • Make sure you're on the master branch with git status, then run $ git pull upstream master. This will perform the fetch and merge operations. Unless you're developing on master (at your own risk) there should not be any merge conflicts.
  4. Make a Branch
    • Make a branch to do your work in. Your branch should be named after the work done on it, like a feature or bug fix.
    • cd to your local copy of your forked repo. Run
      $ git checkout -b Name_of_Branch upstream/master.
  5. Do the Work
    • Do whatever work you want to do. Make an agent with a new algorithm. Do new visualizations or analytics for an old agent. Whatever!
  6. Stage Your Changes
    • Stage your changes using calls to git add and git rm if necessary.
  7. Commit Your Changes
    • Commit your changes with git commit. Go back to step 5 afterwards if you're splitting the work into multiple commits.
  8. Check for Updates One Last Time
    • Run $ git fetch upstream master. This will retrieve any new code from upstream, but it won't really do anything with it.
    • Run $ git rebase upstream/master Name_of_Branch or whatever. This takes your work and sort of re-does it all on top of the most up to date code. That way it is easier to merge into the main project!
  9. Push Your Changes
    • Push your changes to your personal repo with $ git push origin. This will add a branch named Name_of_Branch or whatever to your fork on GitLab. This branch is accessible through the Branches link in the Repository menu on the left side of the main page for your fork on GitLab.
  10. Create a Merge Request
    • If everything looks good to you on the branch then it's time to start merging it into the main repo
    • Visit the GitLab page for your forked repo and click on the + icon below and to the right of the sAnIc icon, then press New merge request. This should take you to a page for merge request creation with boxes for Source branch and Target branch.
    • Under Source branch the Source project on the left should read YourUsername/sAnIc and the Source branch should select Name_of_Branch or whatever. Under Target branch the Target project on the left should read sAnIc-ND/sAnIc and the Target branch should be master. Assuming that is all true, click Compare branches and continue in the bottom left.
    • On this page make sure the title is clear and give a description of the work. Try to be very thorough. The merges need to be reviewed so let the reviewer know all of the information that they might find useful about what you're trying to contribute. Use good formatting too; the description box supports Markdown.
    • Do not assign the merge to anyone
    • You can check Remove source branch when merge request is accepted. if you'd like the branch you pushed to origin deleted from GitLab.
    • Press Submit merge request
  11. Merge Review
    • If upon review your work needs any modifications, perform them on your local machine, stage, commit, update, and push them just as before, but do not make another merge request. The original merge request will be updated with your new commits. The reviewer should help you through this process.
  12. Eventually your work will be merged. You should go back to the local clone of your fork, checkout the master branch, and pull from upstream again (step 3). You're now free to delete Name_of_Branch or whatever with $ git branch -d Name_of_Branch or whatever.

Install the Virtual Environment

  1. Acquire the ROMs from Conrad. We cannot put these in a public repo; they are proprietary. On that note, DO NOT COMMIT/PUSH THESE TO YOUR REPO!! I assume they are stored in sAnIc/roms. If you keep this convention, then that directory will be ignored thanks to .gitignore
  2. Run $ ./ roms. This will check for requirements, create a virtual environment for Python 3 development, install necessary packages, download the retro-contest repository, and install those tools as well.
  3. Run $ source venv/bin/activate. This imports the virtualenv environment into your shell instance and should prepend (venv) to your shell prompt. That means anything installed with pip3.6 will be local to this virtualenv and can be cleanly swept away with rm -r venv. To get out of this environment when you're done working on the project run deactivate, it's that simple.

Start a Jupyter Notebook in the Virtualenv

Jupyter notebooks are good for data analysis because you can re-run portions of code as needed, without having to recreate or save and load large data structures.

IMPORTANT_NOTE: steps 1 to 3 do not need to be repeated. After creating the kernel it can be reused at will.

  1. Install Jupyter for your distribution
  2. Activate the virtualenv (see step 3 here).
  3. Run $ ipython kernel install --user --name=sAnIc. This will create a kernel called sAnIc that we will use to make a notebook in the next step.
  4. Outside of the virualenv (i.e. (venv) should not be prepended to your shell prompt) run $ jupyter notebook. This will start the notebook server and print some text with a link. Copy this link to the address bar in your internet browser.
  5. Navigate to the place you'd like to put the notebook and click 'new' in the upper right corner; this will open a new context. In that context, under 'Notebook:' click 'sAnIc'; this designates the kernel we created earlier, giving you access to the virtualenv's resources.
  6. Use the notebook!

Creating an Agent

First and foremost, look at agents/ for reference on the most basic implementation of an agent.

All agent modules must live directly in the agents folder, and not in any subdirectory, to the packaging necessary for deployment. All agents need to inherit somewhere in their lineage from the Agent class found in This provides the basic functionalities, main implementation, and command line parsing tools. There are 4 basic parts to an agent.

  1. Agent Class Definition and __init__
    • Don't forget the shebang on the first line #!/usr/bin/env python3.6
    • All agent class definitions should look like
      class AgentBeingDefined(AgentOrAgentChild). Multiple inheritance can be useful here (see agents/
    • __init__ may be inherited if the agent being defined does not require any more parameters than its parents. However, if new parameters are being introduced by the agent being defined, then __init__ must be implemented to handle them.
    • __init__ should first take as arguments the necessities for agent creation (is_remote, game, state, max_timesteps, do_render, do_monitor) followed by any parameters necessary for the parent __init__ method, finally ending with parameters specific to the agent being defined.
      • The first line in the body of __init__ should be a call to the parent class __init__ that looks like
        super().__init__(is_remote, game, state, max_timesteps, env_wrapper, do_pause, do_render, do_monitor, any_other_parent_args...) where env_wrapper is some class that maintains useful environment information for the agent being defined (see agents/'s use of HistoriedEnv). The default value (BasicEnv) will suffice in most cases.
      • __init__ should then take care of any initialization procedures specific to the agent being defined.
  2. The play() Method
    • This method may be inherited from a concrete class, but cannot be inherited from the abstract Agent class.
    • It should be an infinite loop that essentially plays the game. Remember that env.reset() must be called once before any calls to env.step(), and env.reset() must be called when done is True.
  3. The init_parser() Method
    • This method may be inherited if no extra parameters are necessary for the agent being defined.
    • The first line in this methods body should be a call to the parent definition that looks like
      parser = super().init_parser("description of class being defined")
    • Finally you should add parameters specific to the agent being defined with calls to parser.add_argument and finally return parser.
  4. main() Invocation
    • You can rely on Agent's main() static method. The bottom of your agent should look like

      if __name__ == '__main__':
        except gre.GymRemoteError as e:
          print('exception', e)
    • This will call the main inherited from Agent with AgentBeingDefined as an argument so that the main() body will instantiate an agent of type AgentBeingDefined and call that instantiation's play() method.

Once all of that is done you can run your agent right from the command line with $ ./ Agent provides some nice flags like --render to watch and control the agent at work and --monitor to produce a monitor.csv. See $ ./ --help for the complete list.

Simulate an Agent Locally

This process will mimic the remote environment used for official scoring when submitting a job. It is important to do this before submitting the agent as a job so that you don't waste time finding runtime errors after uploading to the server.

  1. Activate your virtual environment (see step 3 here).
  2. Acquire a credentials file from Conrad. This contains sensitive passwords and stuff. DO NOT COMMIT/PUSH THIS FILE!! If you keep the name retro_contest_credentials.dontcommit then it will be safely ignored thanks to .gitignore. If you guys are paranoid, then we can work out some encryption, but for now just be careful.
  3. Run $ [--results_dir RESULTS_DIR] [--args ARGS] path name version game state timestep_limit'. This will
    • Create a Docker container with the agent at path installed as a Python package.

    • Tag that container as team_member_name/name:version. The name should characterize the implementation (e.g. DNN for deep neural net) and version is a convenient way to iteratively experiment on an implementation.

    • Simulate a contest run using the new container for the agent and the specified game and state for the environment. It will pass on --args if provided to the agent invocation, like parameter values (--remote is automatically provided).

      The timestep_limit determines how many time steps (i.e. calls to env.step()) the evaluation is alloted. It doesn't matter how many times Sonic dies, wins, or runs out of time; the evaluation only has the allotted amount of steps to achieve the highest average score possible. The contest sets this to one million steps; you'll have to play around with this value to fit your experimentation needs.

  4. This will output a folder named results (by default. see --results-dir) in the working directory. In this directory there are useful files:
    • agent_stderr.txt and agent_stdout.txt: This is where you'll find any debugging output or errors produced by your agent.
    • remote-stderr.txt and remote-stdout.txt: This is where you'll find output and errors from the remote environment provided by the retro-contest people.
    • log.csv: This is just a recording of the time elapsed for every 1000 steps. It should give you an idea of how fast your agent is running.
    • monitor.csv: Every row represents a trial the evaluation performed. The column r is the reward for that trial, l is the number of steps it took to complete the trial, and t is the wall-clock time it took to complete the trial. This file can be automatically analyzed.
    • bk2 is a directory containing visual information for each trial that can be converted into .mp4 videos. Due to the nature of Docker, this file is recursively owned by root, so manipulating or removing this file will require root permissions, but there is nothing inherently root-worthy in this directory.

Analyze Results and $ ./ --monitor produce a file called monitor.csv (see step 5 here for full details). This file is the key to statistical and graphical analysis. Run $ monitor_path "Title of Experiment" to produce analytical output. This will leave four files in the directory containing monitor_path.

  • stats.json: Obviously this is a serialized JSON object. In it you will find descriptive statistics like mean, median, stddev, etc. regarding rewards and timestep lengths. This can be ingested with json.load().
  • rewards.svg: This is an svg depicting two super-imposed plots: a scatter plot of the rewards of every episode, and a line plot of the mean reward at any given timestep. This can be opened in a web browser.
  • histogram.svg: This is a 10-bin histogram of all rewards. This can be opened in a web browser.
  • analysis.html: This is a webpage displaying the experiment title, stats.json as a table, rewards.svg, and histogram.svg.

Evaluate an Agent On Many Tests

Evaluating an agent on one game locally can be great for debugging, but it's not really what we need for analytics. In the classic data science framework we need to run our agents on test sets, and here's how.

  1. Run $ [--path PATH] [--results_dir RESULTS_DIR] [--nprocs NPROCS] tests_file timestep_limit name version. Every line of the tests_file is an environment to test the agent in; a game title followed by a state name, separated by white space (here's an example). The timestep_limit specifies the timestep limit for each environment. The --nprocs flag allows us to perform these experiments in parallel! And --path can be used if a Docker container needs to be built.
  2. This will output a folder named results (can override with --results_dir). This is similar to singular, local evaluation results, but inside this directory is many more directories. Each subdirectory holds the output of the experiment corresponding to the subdirectory name.

Analyze a Batch of Results

The first thing you might want to do is merge many monitor.csv files into a single, large monitor.csv file. For example, if you ran 30 trials of an agent on the same level with the same parameters, then you are probably interested in the aggregate performance of this agent and not interested in the trials individually. To do this run $ path1 path2... where pathX is a path to some monitor.csv. By default the result is printed, but the -o flag can be used to designate an output filename. The actual merge process is too complicated for this document, but the output has the properties of a single agent performing all of the actions that the constituent agents performed.

Much like analyzing a single result, analyzing a batch of monitor.csv files is done with a script, Run $ output_dir comparison_name name1 path1 name2 path2... to produce results. The arguments are as follows

  • output_dir: Location where results will be stored.
  • comparison_name: Used to title the figures and pages generated. It should characterize the intent behind comparing these results.
  • name1 path1 name2 path2...: This is a list of experiment names and their corresponding monitor.csvs. The experiment names should characterize the data in their monitor.csv in the context of the comparison. These names will be used for legend labels in the figures and pages generated.

This will produce a histogram.svg, rewards.svg, and analysis.html in output_dir that correspond to their single result counterparts, but compiles the data from all of the monitor.csvs under consideration.

Submit a Job

The leaderboard only ranks the most recent result for your team, so this should only be done after careful testing and analysis. Ask Conrad to help you take an agent through these stages.

Convert .bk2 Files to .mp4 Videos

These steps require the ffmpeg application with x264 support.

  1. Make sure the .bk2 file is not owned by root. This can be done to produced results with $ sudo chown -R username:username results/bk2 (this is an example, your invocation may vary).
  2. Activate your virtual environment (see step 3 here).
  3. To store the output use $ python3 -m retro.scripts.playback_movie somefile.bk2. This will produce somefile.mp4 in the working directory. To immediately view the output use python3 -m retro.scripts.playback_movie somefile.bk2 -v viewer where viewer is the name of the application you'd like to use to view the movie (e.g. vlc, mplayer, etc).


Contest Websites

  • Poster/Front Page: The basics of the contest.
  • Blog Post: Some details about the work they've already put in and their hopes for the contest.
  • Details Page: Digs into the meat of the contest and the tools that they provide.
  • Leaderboard: The current, live contest results.
  • Benchmark Whitepaper: This is an academic paper describing the benchmark they use for measuring performance.
  • Discord Server: Go here to reach out to the community and admins for questions or the Troubleshooting channel. In my experience they're very helpful, as long as it's not the weekend.


NOTE: Visit these links and peruse the READMEs; they often have very valuable information.

  • Gym: A toolkit for developing and comparing reinforcement learning algorithms. This is the kernel of the contest.
  • Gym Retro: The foundation of the contest. This turns emulators into Gym environments. It's what allows us to interact with the Sega Genesis video games.
  • Retro-Contest: Contest-specific wrappers for Gym-Retro and various support tools.
  • Retro-Baselines: Contest entry examples that the OpenAI team has pre-built. You can try these out for yourself with a little modification.