Learning Montezuma’s Revenge from a single demonstration
OpenAI just cracked Montezuma's Revenge from a single demo—a benchmark that stumped the field for years.
Why it matters
OpenAI demonstrates a major breakthrough in sample efficiency and reinforcement learning, showing agents can learn complex, long-horizon tasks from minimal human guidance. This is directly relevant to how AI systems will learn to perform real-world tasks at scale.
The key facts
11 to knowScore of 74,500 on Montezuma's Revenge from single demonstration
Outperforms all previously published results
Algorithm uses PPO (Proximal Policy Optimization)
Agent learns by playing sequences from carefully chosen states
Same RL algorithm that powers OpenAI Five
Published July 4, 2018
Agent achieved 74,500 score on Montezuma's Revenge from single human demonstration
Better than any previously published result at time of publication
Used PPO (Proximal Policy Optimization) reinforcement learning algorithm
Algorithm learns by playing sequences from carefully chosen demonstration states
Published July 4, 2018 by OpenAI
Go to the source
OpenAI Blogopenai.com
Publisher excerpt: We’ve trained an agent to achieve a high score of 74,500 on Montezuma’s Revenge from a single human demonstration, better than any previously published result. Our algorithm is simple: the agent plays a sequence of games starting from carefully chosen states from the demonstration, and learns from…