Artificial intelligence · 2017

RL from human preferences

Train a reward model from human comparisons, then optimise against it.

Paul Christiano, Jan Leike and colleagues at OpenAI and DeepMind taught agents tasks like backflips from about 900 bits of human feedback. Reinforcement learning from human feedback became the main technique for aligning language models with what people want.

See it in the living treeWikipedia