pyrddlgym-project/rddlrepository

Hosts domain and instance RDDL files, covering problems from a wide range of disciplines, integration with the pyRDDLGym ecosystem.

rddlrepository

Purpose:

• contains a growing collection of standalone problems, and archived problems from previous International Planning Competitions
• domains are selected across a variety of disciplines, and showcase various aspects of the latest official RDDL syntax, to serve as starting points for learning the language and designing custom domains
• contains custom visualizers for a subset of domains, to be used with the pyRDDLGym package
• out-of-the-box compatibility with the pyRDDLGym ecosystem

Contents

• Installation
• Usage
  • Domains
  • Instances
  • Loading the Environment in pyRDDLGym
  • Registering your Own Problems
• Domains Included

Installation

We require Python 3.8+ and numpy. To install via pip:

pip install rddlrepository

To install via github

git clone https://github.com/pyRDDLGym-project/rddlrepository.git

To make use of the existing visualizers, you also require pyRDDLGym (pip install pyRDDLGym).

Usage

Domains

The core object for extracting domains and their instances is the RDDLRepoManager object:

from rddlrepository.core.manager import RDDLRepoManager
manager = RDDLRepoManager(rebuild=True)

Note

rebuild instructs the manager to rebuild the manifest, which is an index containing the locations of all domains and instances for fast access. While you do not need this option in normal operation, in case you add your own domains or the manifest becomes corrupt, you can force it to be recreated.

To list all domains in rddlrepository:

print(manager.list_problems())

Problems are organized by context (e.g. year of the competition, standalone):

print(manager.list_context())
print(manager.list_problems_by_context("standalone"))   # list all standalone problems
print(manager.list_problems_by_context("ippc2018"))     # list all problems from IPPC 2018

Instances

The information for a specific domain is a ProblemInfo instance:

problem_info = manager.get_problem("EarthObservation_ippc2018")

To list all the instances of a domain:

print(problem_info.list_instances())

To return the paths of the domain and instance (1)

print(problem_info.get_domain())
print(problem_info.get_instance("1"))

To return the pyRDDLGym visualizer class:

viz_class = problem_info.get_visualizer()

Loading the Environment in pyRDDLGym

This information can be used in pyRDDLGym to instantiate an OpenAI Gym environment:

import pyRDDLGym
env = pyRDDLGym.make(domain=problem_info.get_domain(), instance=problem_info.get_instance("1"))
env.set_visualizer(problem_info.get_visualizer())

However, a shorter way is:

import pyRDDLGym
env = pyRDDLGym.make("EarthObservation_ippc2018", "1")

In either case, env is an instance of OpenAI Gym's Env class with the reset() and step() function already implemented for you!

Registering your Own Problems

To register a new context in rddlrepository for later access:

manager.register_context("MyContext")

To register a new problem in a given context for later access:

domain_content = """
    domain ... {
    ...
    }
"""
manager.register_domain("MyDomain", "MyContext", domain_content,
                        desc="a description of this domain", viz="ModuleName.ClassName") 

Here, "ModuleName.ClassName" refers to the Module name and the Class name of the visualizer (optional).

To register an instance for an existing domain for later access:

instance_content = """
    instance ... {
    ...
    }
"""
problem_info.register_instance("MyInstance", instance_content)

Domains Included

A downloadable PDF file listing all domains currently supported in rddlrepository can be found here.

Domains are broken down into contexts, which includes some standalone domains, as well as domains used in the past competitions:

• IPC 2011: http://users.cecs.anu.edu.au/~ssanner/IPPC_2011/
• IPC 2014: https://ssanner.github.io/IPPC_2014/
• IPC 2018: https://ipc2018-probabilistic.bitbucket.io/
• IPC 2023: https://ipc2023.github.io.

rddlrepository's Issues

Add documentation, i.e. readthedocs integration, update readme, once v2 is merged with main (Difficulty: EASY)

Allow all 8 domains in IPPC 2023 to be indexable, fix standalone domains choice, integrate viz better

I think there is a conflict of names between IPPC 2023 and standalone (i.e. two versions of the UAV continuous problem even though both domain.rddl are the same, maybe we need to remove one from mixed/discrete and make the other standalone?). Viz from standalone should be distributed anywhere it is relevant to past IPC (e.g. does elevator viz work for Elevator 2011?)

More error handling and robustness (Difficulty: EASY)

Raise exception when loading domains with identical names, don't know what other cases to catch (Difficulty: EASY)

Write script to auto compile table (e.g. latex?) listing all domains (and link in readme instead of ugly tables)