Reward Reports for Reinforcement Learning Towards Documentation and Understanding of Dynamic Machine Learning Systems

‼️ HIRING! We are hiring an intern(s) to help build out Reward Reports! If interested in ethical AI, documentation, or sociotechnical considerations of RL systems, please contact thomaskrendlgilbert at gmail dot com.

---

Building Accountable and Transparent RL

This RLDM 2022 Workshop is on Saturday June 11th from 1:00-5:00pm EST.

You can find the slides for the workshop here.

Motivation

When RL is used in societally relevant domains, practitioners must balance contested values, while anticipating and responding to resulting downstream effects. This will require documenting the behavior of RL systems over time, both in relation to design choices and dynamic effects. In this workshop, we will survey existing approaches to AI documentation, discuss the unique challenges posed by RL deployments, and introduce a proposal for “Reward Reports”: living documents that demarcate design choices and track system behavior. The majority of the workshop will be interactive, with participants trying out and critiquing methods of documentation for real or imagined RL applications.

To learn more about Reward Reports, see the Reward Reports paper, the CLTC RL Risks Whitepaper, or the github repo with a template.

Format – The (Un)Workshop

We intend for this workshop to be a generative and interactive space for participants to try out and critique our Reward Reports proposal. As such, activities will be centered around doing and sharing in small groups, punctuated by periodic larger group discussions. We hope that by the end of the workshop, there will be two concrete outputs: a rich set of example RRs, and a list of refinements for the Reward Reports framework.

Agenda

• 1-1:45pm Introduction
  • Welcome and ice breaker
  • Introduction to documentation and lessons learned from deployed systems
• 1:45-4pm Participatory deep dive on real-world applications
  • Brainstorm dynamic harms
  • Interactive documentation exercise
  • Break
  • Articulate an oversight agenda for RL systems (e.g. changelog, impact assessments)
• 4-5pm Reflections and agenda setting
  • Identify missing documentation components / features
  • Propose application-specific protocols
  • Interpret documentation in the context of RL / AI governance

How to participate

Register for the Multi-disciplinary Conference on Reinforcement Learning and Decision Making (RLDM).

Send us a message with any questions!