OpenTTDLab - Run reproducible experiments using OpenTTD

OpenTTDLab is a Python framework for using OpenTTD to run reproducible experiments and extracting results from them, with as few manual steps as possible. OpenTTDLab can also be used to help run regression tests of OpenTTD AIs, parse OpenTTD savegame files, and download content from BaNaNaS.

OpenTTDLab is based on Patric Stout's OpenTTD Savegame Reader.

Caution

OpenTTDLab currently does not work with OpenTTD 14.0 or later. The latest version of OpenTTD known to work is 13.4.

---

Contents

• Features
• Installation
• Running experiments
• Plotting results
• Examples
• API
  • Core function
  • Configuring AIs
  • Configuring AI libraries
  • Parsing savegame files
  • Downloading from BaNaNaS
• Tips for repeatability, reproducibility, and replicability
• Compatibility
• Licenses and attributions

---

Features

OpenTTDLab essentially turns OpenTTD into a simulator - and through AIs and AI libraries it allows you to experiment with different techniques of building supply chains and study their effects. In more detail OpenTTDLab:

• Allows you to easily run OpenTTD in a headless mode (i.e. without a graphical interface) over a variety of configurations.
• And allows you to do this from Python code - for example from a Jupyter Notebook.
• As is typical from Python code, it is cross platform - allowing to share code snippets between macOS, Windows, and Linux, even though details like how to install and start OpenTTD are different on each platform.
• Downloads (and caches) OpenTTD, OpenGFX, AIs, and AI libraries - no need to download these separately or through OpenTTD's built-in content browser.
• Transparently parallelises runs of OpenTTD, by default up to the number of CPUs. (Although with fairly poor scaling properties.)
• Results are extracted from OpenTTD savegames as plain Python dictionaries and lists - reasonably convenient for importing into tools such as pandas for analysis or visualisation.

Installation

OpenTTDLab is distributed via PyPI, and so can usually be installed using pip.

python -m pip install OpenTTDLab

When run on macOS, OpenTTDLab has a dependency that pip does not install: 7-zip. To install 7-zip, first install Homebrew, and then use Homebrew to install the p7zip package that contains 7-zip.

brew install p7zip

You do not need to separately download or install OpenTTD (or OpenGFX) in order to use OpenTTDLab. OpenTTDLab itself handles downloading them.

Running experiments

The core function of OpenTTD is the run_experiments function.

from openttdlab import run_experiments, bananas_ai

# Run experiments...
results = run_experiments(
    openttd_version='13.4',  # ... for a specific versions of OpenTTD
    opengfx_version='7.1',   # ... and a specific versions of OpenGFX
    experiments=(
       {
         # ... for random seeds
         'seed': seed,
         # ... running specific AIs. In this case a single AI, with no
         # parameters, fetching it from https://bananas.openttd.org/package/ai
         'ais': (
             bananas_ai('54524149', 'trAIns', ai_params=()),
         ),
         # ... each for a number of (in game) days
         'days': 365 * 4 + 1,
       }
       for seed in range(0, 10)
    ),
)

Plotting results

OpenTTD does not require any particular library for plotting results. However, pandas and Plotly Express are common options for plotting from Python. For example if you have a results object from run_experiments as in the above example, the following code

import pandas as pd
import plotly.express as px

df = pd.DataFrame(
    {
        'seed': row['experiment']['seed'],
        'date': row['date'],
        'money': row['chunks']['PLYR']['0']['money'],
    }
    for row in results
)
df = df.pivot(index='date', columns='seed', values='money')
fig = px.line(df)
fig.show()

should output a plot much like this one.

Examples

A few examples are available:

• A Jupyter notebook of the above example, briefly exploring the performance of trAIns
• A Jupyter notebook showing how the performance of OpenTTDLab scales with the number of workers
• A pytest .py file using OpenTTDLab as a test harness for an OpenTTD AI, run automatically using a GitHub action workflow

API

Core function

run_experiments(...)

The core function of OpenTTDLab is the run_experiments function, used to run an experiment and return results extracted from the savegame files that OpenTTD produces. It has the following parameters and defaults.

• experiments=()

  An iterable of the experiments to run. Each experiment should be a dictionary with the (string) keys:

  • 'ais'

    The list of AIs to run in this experiment. See the Fetching AIs section for details on this parameter.

  • 'seed'

    The integer seed of the random number generator for this experiment.

  • 'days'

    The integer number of in-game days that this experiment will run for.

  • openttd_config=''

    OpenTTD config to run each experiment under. This must be in the openttd.cfg format. This is added to by OpenTTDLab before being passed to OpenTTD.

• ais_libraries=()

  The list of AI libraries to have available to AI code. See the Fetching AI libraries section for details on this parameter.

• result_processor=lambda r: (r,)

  A function that takes a single result row, which is a parsed save game file from an experiment, alongside other metadata describing the experiment, and returns it processed in some way. The function should return an iterable of zero or more rows that will appear in the the return value of run_experiments.

  This is typically used to reduce memory usage with high numbers of experiments where only a small amount of data is needed for analysis.

• final_screenshot_directory=None

  The directory to save a PNG screenshot of the entire map at the end of each run. Each is named in the format <seed>.png, where <seed> is the experiment's seed of the random number generator. If None, then no screenshots are saved.

  For technical reasons, a window will briefly appear while each screenshot is being saved. This can be avoided when running on Linux if xvfb-run is installed and available in the path.

• max_workers=None

  The maximum number of workers to use to run OpenTTD in parallel. IfNone, then os.cpu_count() defined how many workers run.

• openttd_version=None

  The version of OpenTTD to use. If None, the latest version available at openttd_base_url is used.

  Caution OpenTTDLab currently does not work with OpenTTD 14.0 or later. The latest version of OpenTTD known to work is 13.4.

• opengfx_version=None

  The version of OpenGFX to use. If None, the latest version available at opengfx_base_url is used.

• openttd_base_url='https://cdn.openttd.org/openttd-releases/

  The base URL used to fetch the list of OpenTTD versions, and OpenTTD binaries.

• opengfx_base_url='https://cdn.openttd.org/opengfx-releases/

  The URL used to fetch the list of OpenGFX versions, and OpenGFX binaries.

• get_http_client=lambda: httpx.Client(transport=httpx.HTTPTransport(retries=3)

  The HTTP client used to make HTTP requests when fetching OpenTTD, OpenGFX, or AIs. Note that the bananas_ai function uses a raw TCP connection in addition to HTTP requests, and so not all outgoing connections use the client specified by this.

Configuring AIs

The value of the ais key of each dictionary in the experiments parameter configures which AIs will run, how their code will be located, their names, and what parameters will be passed to each of them when they start. In more detail, the ais parameter must be an iterable of the return value of any of the the following 4 functions.

Important

The ai_name argument passed to each of the following functions must exactly match the name of the corresponding AI as published. If it does not match, the AI will not be started.

Important

The return value of each of the following is opaque: it should not be used in client code, other than by passing into run_experiments as part of the ais parameter.

bananas_ai(unique_id, ai_name, ai_params=(), md5=None)

Defines an AI by the unique_id and ai_name of an AI published through OpenTTD's content service at https://bananas.openttd.org/package/ai. This allows you to quickly run OpenTTDLab with a published AI. The ai_params parameter is an optional parameter of an iterable of (key, value) parameters passed to the AI on startup.

The unique_id is sometimes surfaced as the "Content Id", but it should not include its ai/ prefix.

If you pass the full MD5 hex string of a specific version of AI as md5, for example previously returned from the download_from_bananas function, the corresponding version will be used. Otherwise, the latest version will be used.

local_folder(folder_path, ai_name, ai_params=()))

Defines an AI by the folder_path to a local folder that contains the AI code of an AI with name ai_name. The ai_params parameter is an optional parameter of an iterable of (key, value) parameters passed to the AI on startup.

local_file(path, ai_name, ai_params=())

Defines an AI by the local path to a .tar AI file that contains the AI code. The ai_params parameter is an optional parameter of an iterable of (key, value) parameters passed to the AI on startup.

remote_file(url, ai_name, ai_params=())

Fetches the AI by the URL of a tar.gz file that contains the AI code. For example, a specific GitHub tag of a repository that contains its code. The ai_params parameter is an optional parameter of an iterable of (key, value) parameters passed to the AI on startup.

Configuring AI libraries

The ai_libraries parameter of run_experiments ensures that AI libraries are available to the AIs running. In more detail, the ais_libraries parameter must be an iterable, where each item the the return value of the bananas_ai_library function described below.

Note that for AIs specified by bananas_ai OpenTTDLab automatically downloads all of their AI library dependencies without them having to be specified through the ai_libraries parameter. This includes all transitive AI library dependencies - AI libraries needed by AI libraries needed by AIs, and so on.

Similarly for AI libraries specified by bananas_ai_library - OpenTTDLab automatically downloads of their AI library dependencies.

bananas_ai_library(unique_id, ai_library_name, md5=None)

Fetches the AI library defined by unique_id and ai_name of a library published through OpenTTD's content service at https://bananas.openttd.org/package/ai-library.

The unique_id is sometimes surfaced as the "Content Id", but it should not include its ai-library/ prefix.

If you pass the full MD5 string of a specific version of AI library as md5, for example previously returned from the download_from_bananas function, this will fetch this corresponding version from BaNaNaS. Otherwise, the latest version is fetched.

Parsing savegame files

parse_savegame(chunks: Iterable[bytes])

Under the hood the run_experiments handles the generation and parsing of savegame files, but if you have your own savegame files generated separately, the parse_savegame function is exposed that can extract data from them.

It takes an iterable of bytes instances of a savegame file, and returns a nested dictionary of parsed data.

from openttdlab import parse_savegame

with open('my.sav') as f:
   parsed_savegame = parse_savegame(iter(lambda: f.read(65536), b''))

Downloading from BaNaNaS

Important

Please do not use this to try to download all content from BaNaNaS. See this discussion about writing a client for BaNaNaS for more details.

Important

Please note the license of each piece of content you download, and adhere to its rules. As examples, licenses may require you to attribute the author, they can restrict you from distributing any modifications you make, they can restrict you from using the content for commercial purposes, or they can require you to make the source available if you distribute a compiled version.

download_from_bananas(content_id: str, md5: Optional[str]=None)

This function is a Python BaNaNaS client for downloading the content from BaNaNaS. Given a content id, it returns an iterable of that content and all of its direct and transitive dependencies.

from openttdlab import download_from_bananas

with download_from_bananas('ai/41444d4c') as files:
    for content_id, filename, license, partial_or_full_md5, get_data in files:
        with get_data() as chunks:
            with open(filename, 'wb') as f:
                for chunk in chunks:
                    f.write(chunk)

Each chunks iterable are the binary chunks of the non-compressed .tar file of the content. Also, under the hood download_from_bananas transparently caches content where possible. This is the main reason for using context managers as in the above example - they allow for robust cleanup of resources and caching of data once the data has been iterated over.

If you don't pass md5, download_from_bananas will return details for the latest version of the content. And if the content has a known and acceptable license, partial_or_full_md5 will contain the full MD5 for the content, which can then be subsequently passed back into download_from_bananas to download the same version later. If the file does not have a known license, download_from_bananas will contain the only the first 8 characters of the MD5. This means that download_from_bananas is deterministic only for content that has an acceptable license, and if you pass an MD5 previously retrieved from a call to download_from_bananas.

Note that the function run_experiments that uses bananas_ai or bananas_ai_library will handle automatically downloading from BaNaNaS, so this function is usually only useful if you would like to run experiments without using the bananas_* functions, or report on the filename (which includes the version of each piece of content) or the MD5 sum of the file.

Tips for repeatability, reproducibility, and replicability

OpenTTDLab should be able to help with all the 3Rs of repeatability, reproducibility and (although to a lesser degree) replicability of simulations. Definitions of the 3Rs are taken from https://www.acm.org/publications/policies/artifact-review-and-badging-current.

Repeatability

"For computational experiments, this means that a researcher can reliably repeat her own computation."

The run_experiments function is the primary way OpenTTLab supports repeatability - with a single function call you can run a range of experiments. But to make sure you get the same results on each invocation:

• Pin to a specific OpenTTD and OpenGFX version.
• If fetching AIs and AI libraries from BaNaNaS, use the MD5 to ensure the same versions are used (at the moment the only way of discovering these is to initially separately call the download_from_bananas_function).
• If fetching AIs and AI libraries from another remote source, make sure that source is immutable.
• For all your own code store in version control, e.g. git, and make sure to note version/commit IDs used to generate results.
• Use a virtual environment to pin Python version and all dependencies (most importantly of OpenTTDLab).
• Use fixed random seeds.
• (Out of paranoia mostly, note the OS and its version, and CPU type.)

Reproducibility

"For computational experiments, this means that an independent group can obtain the same result using the author’s own artifacts."

To help others reproduce your results:

• Take all the repeatability steps above, and make sure to communicate all the noted details, e.g. what versions where pinned, along with any results.
• Make any artifacts that you have created available alongide the results / linked from the results. This can include the source of AIs you have used or created, the Python code to run OpenTTDLab, and the Python code to analyse the results.

In general, bitwise reproducibility is not expected, but with fixed random seeds it "should" be the case with OpenTTD/OpenTTDLab (depending on the analysis performed).

Replicability

"For computational experiments, this means that an independent group can obtain the same result using artifacts which they develop completely independently."

This is arguably the most difficult of the 3Rs to plan for, and for other researchers to achieve. To help others replicate your results:

• Take all the repeatability and reproducibility steps above.

• Note and communicate the high level algorithms involved, so others can create their own implementations of them. For example, if you have written an AI, then precisely communicate the algorithm that AI takes so others can implement it, and then run it with OpenTTDLab.

  For full replicability, it could be argued that OpenTTD itself, or something equivalent, would also have to be implemented. While this sounds unfeasbile, depending on what the AIs involve and what you argue the results are, it might be feasible. However, for this level of replicability OpenTTDLab is unlikely to be helpful.

Compatibility

• OpenTTD versions between 12.0 and 13.4 (OpenTTD >= 14.0 is not currently supported. See this discussion on the changes in OpenTTD 14.0.)
• Linux (tested on Ubuntu 20.04), Windows (tested on Windows Server 2019), or macOS (tested on macOS 11)
• Python >= 3.8.0 (tested on 3.8.0 and 3.12.0)

Licenses and attributions

TL;DR

OpenTTDLab is licensed under the GNU General Public License version 2.0.

In more detail

OpenTTDLab is based on Patric Stout's OpenTTD Savegame Reader, licensed under the GNU General Public License version 2.0.

The OpenTTDLab logo is a modified version of the OpenTTD logo, authored by the OpenTTD team. The OpenTTD logo is also licensed under the GNU General Public License version 2.0.

The .gitignore file is based on GitHub's Python .gitignore file. This was originally supplied under CC0 1.0 Universal. However, as part of OpenTTDLab it is licensed under GNU General Public License version 2.0.

trAIns is authored by Luis Henrique O. Rios, and licensed under the GNU General Public License version 2.0.

OpenTTD and OpenGFX are authored by the OpenTTD team. Both are licensed under the GNU General Public License version 2.0.