Australian Responsible Autonomous Agents Collective (araac.au)
ActiveResearch focus
Researching the ways in which multi-objective approaches to reward specification and reinforcement learning can improve the safety/alignment and explainability of agents. See for example the following papers:
Human-aligned artificial intelligence is a multiobjective problem https://dl.acm.org/doi/abs/10.1007/s10676-017-9440-6
Scalar reward is not enough: A response to Silver, Singh, Precup and Sutton (2021) https://link.springer.com/article/10.1007/s10458-022-09575-5
Potential-based multiobjective reinforcement learning approaches to low-impact agents for AI safety https://www.sciencedirect.com/science/article/pii/S0952197621000336
Intent-aligned AI Systems Must Optimize for Agency Preservation https://openreview.net/pdf?id=rfvgdfd1K9
PhD students (particularly Australian citizens, as options for funded scholarships for international applicants are limited).
Collaboration with researchers from other organisations.
Support with funding.
Damaqu Fireside
Papers on Damaqu (4)
- Multi-objective Reinforcement Learning: A Tool for Pluralistic Alignmentarxiv· 15 Oct 2024
- Scalar reward is not enough: a response to Silver, Singh, Precup and Sutton (2021)other· 16 Jul 2022
- A Practical Guide to Multi-Objective Reinforcement Learning and Planningarxiv· 17 Mar 2021
- Human-aligned artificial intelligence is a multiobjective problemother· 4 Oct 2017