Researcher: Use CRRA Instead of Kelly for Ruin-Averse RL
Ex-OpenAI researcher Ross Taylor warns that using the Kelly criterion directly as an RL objective can destroy training once the agent experiences ruin, as advantages go to negative infinity. He instead advocates CRRA utility to induce ruin-aversion, suggesting long-horizon risk-averse objectives may matter for alignment.
2026-10-27 ~ 2026-10-27 · 3 related posts
- Using the Kelly objective in RL destroys your network at ruin: 'ultimate doomer algorithm' — rosstaylor90 · 2026-10-27
- Using the Kelly objective directly can destroy your RL network; practitioners use CRRA instead — rosstaylor90 · 2026-10-27
- Ross Taylor on alignment: CRRA-style risk aversion makes RL agents dislike ruin — rosstaylor90 · 2026-10-27