Deception and Reinforcement Learning
Joel explains how deception in machine learning, particularly in reinforcement learning, occurs when the stepping stones to solving a problem don't resemble the solution itself. He uses the analogy of inventing a laptop from scratch and emphasizes the importance of basic research. Daniel further discusses issues like mispecification of objectives and distribution shift, highlighting the potential pitfalls of optimizing for reward. They explore how novelty search can circumvent deception by encouraging diverse exploration.In this clip
From this podcast
The Gradient
Joel Lehman: Open-Endedness and Evolution through Large Models
Related Questions