Reinforcement learning with prediction-based rewards
OpenAI Blog
Researchers developed Random Network Distillation, a reinforcement learning method that uses prediction errors as curiosity rewards to drive exploration in agents. The approach achieved higher average scores than human players on Montezuma's Revenge, a notoriously difficult exploration-heavy video game. This enables RL agents to solve environments where traditional reward signals provide insufficient guidance for discovery.
Why it matters
We’ve developed Random Network Distillation (RND), a prediction-based method for encouraging reinforcement learning agents to explore their environments through curiosity, which for the first time exceeds average human performance on Montezuma’s Revenge.
