OpenTTDLab 0.0.58

Python framework for running reproducible experiments using OpenTTD

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 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 an experiment
• Plotting results
• Examples
• API
• Compatibility
• Licenses and attributions

---

Features

• 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 notebook of the above example and an example measuring the performance of OpenTTDLab are in the examples folder.

API

Running experiments

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.

Fetching 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=())

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.

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.

Fetching 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)

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.

Compatibility

• 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.