In this workshop we introduce Regym, a Deep (multi-agent) RL research framework developed by PhD candidates at York (your humble contributors).

Research in Deep Reinforcement Learning (DRL) is notoriously hard due to all of the moving pieces involved in an experiment. Practitioners need to first make sure that the environment they work with features a clean interface which permits DRL agents to interact with it, which oftentimes is environment specific. Only when this is done, work on the development of the DRL algorithm may begin. This non-trivial process means that most online implementations of the theoretically general-purpose DRL algorithms used in research are seldom programmed to be reused. Hence, DRL researchers spend weeks implementing other researchers’ algorithms to benchmark their own contributions, adding a gargantuan task to their own research.

Regym was built with an emphasis on cross-compatibility across environments and clear interfaces to introduce new algorithms and compare them against existing ones. Regym features readily available implementations of many state of the art DRL algorithms of the last five years. The framework currently supports OpenAI gym and Unity environments. For IGGI PhD candidates, having a shared codebase of RL algorithms will strengthen the relationship between RL researchers across all the member universities.

We sincerely hope that Regym will grant IGGI researchers with a tool to drastically reduce the implementation time cost of their future publications by

1. removing the need to implement an interface between their environments and their DRL algorithm and,
2. giving them access to a pool of already implemented algorithms which they can use for benchmarking purposes out-of-the-box. And we can proudly say that some recent IGGI RL publications have already been implemented using Regym.

This framework has recently become mature enough for public use. By giving a workshop on how to use it we hope to start spreading the word about its existence and, most importantly, get feedback from RL researchers on what parts of the framework could be improved.

Requirements for attendees:

• Basic knowledge of Python
• Basic knowledge of Reinforcement Learning or interest on machine learning work-flows and development tools
• We will ask the attendees to download and install our framework (a Python package)
  

Workshop Organisers: Daniel Hernandez and Kevin Denamganaï