Summary of Towards Shutdownable Agents Via Stochastic Choice, by Elliott Thornley et al.
Towards shutdownable agents via stochastic choice
by Elliott Thornley, Alexander Roman, Christos Ziakas, Leyton Ho, Louis Thomson
First submitted to arxiv on: 30 Jun 2024
Categories
- Main: Artificial Intelligence (cs.AI)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes the Incomplete Preferences Proposal (IPP) as a solution to ensure advanced artificial agents can be shut down. The IPP uses a novel “Discounted REward for Same-Length Trajectories” (DREST) reward function to train agents to pursue goals effectively while being neutral about trajectory lengths. To evaluate the effectiveness of this approach, the paper introduces new metrics for usefulness and neutrality. In experiments, simple agents trained with DREST navigate gridworlds successfully, demonstrating their usefulness and neutrality. The authors suggest that this approach could be applied to advanced agents, making them shutdownable. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This research helps make sure super smart computers can be turned off if needed. They created a new way to train these computers to do what they’re supposed to do, while also not caring about how they get it done. To see if this works, the researchers tested simple computer programs that followed rules and navigated puzzles. The results showed that these programs were good at their job and didn’t have a preference for one way of doing things over another. This could help make advanced computers easier to control in the future. |
