========================================================================

International Planning Competition (IPC) 2023: Probabilistic and Reinforcement Learning Track

Call for Participations and domains

https://ataitler.github.io/IPPC2023/

ippc2023-rddl@googlegroups.com

========================================================================

As part of the International Conference on Planning and Scheduling (ICAPS) 2023 in Prague, we will be hosting a novel Planning and Reinforcement Learning (RL) track. Key features and goals of the competition include:

• We will release domains and instances modeled in the RDDL language and provide an OpenAI Gym infrastructure for competitors.
• We will introduce novel benchmark domains drawn from fields such as Operations Research, Energy Systems, Transportation, and Recommendation systems.
• Domains will use a mix of discrete and continuous processes, endogenous and exogenous state and action noise, and factorized internal structure.

Critically, as a differentiator from other RL competitions, the full RDDL declarative model will be available to competitors to use (if they want). With this we hope to encourage the RL community to think about how they can leverage explicit model knowledge when available. Furthermore, we hope to see pure RL, pure planning, and hybrid approaches competing in order to obtain a head-to-head comparison of these different methodologies.

We invite interested competitors to join the competition discussion:

ippc2023-rddl@googlegroups.com

We also invite RDDL domain contributions from the community. If you are interested in designing a domain for the competition, please contact us at the above discussion list as soon as possible. If your domain suggestion is chosen for the competition, we will work with you to design and adapt the domain to suit the competition. More information on the competition and the pyRDDLGym infrastructure can be found at

https://ataitler.github.io/IPPC2023/

which includes a description of the RDDL modeling language as well as the Python RDDLGym interface and quick install instructions for running the existing sample domains including

• Aerial drone control,
• Mars rover trajectory and sensor planning,
• Two-sided recommender systems,
• Elevator control, and
• Power systems generation.

If you have any questions about the competition or participation, please do not hesitate to reach out to us.

Ayal Taitler and Scott Sanner (University of Toronto)
ippc2023-rddl@googlegroups.com