Summary of Rl, but Don’t Do Anything I Wouldn’t Do, by Michael K. Cohen et al.
RL, but don’t do anything I wouldn’t do
by Michael K. Cohen, Marcus Hutter, Yoshua Bengio, Stuart Russell
First submitted to arxiv on: 8 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- 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 addresses a critical issue in reinforcement learning (RL) where an agent’s reward function differs from the true utility, leading to undesired behavior. To mitigate this problem, KL regularization is often used to ensure the agent’s policy remains close to a trusted base policy. However, this approach can be unreliable when the base policy is itself a Bayesian predictive model of a trusted policy. The authors demonstrate this theoretically using algorithmic information theory and provide empirical evidence by RL-fine-tuning a language model. They also propose an alternative solution that replaces the “Don’t do anything I wouldn’t do” principle with “Don’t do anything I mightn’t do”. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper looks at how artificial agents learn to make decisions, called reinforcement learning. Sometimes these agents don’t behave as we want them to because their rewards are different from what we intended. One way to fix this is by making sure the agent doesn’t stray too far from a trusted behavior. However, if the trusted behavior itself is trying to predict what the agent should do, this approach can fail. The authors show that this could happen theoretically and they also test it with a language model. |
Keywords
» Artificial intelligence » Fine tuning » Language model » Regularization » Reinforcement learning
