Evolution strategies as a scalable alternative to reinforcement learning | OpenAI

We use cookies and similar technologies to deliver, maintain, improve our services and for security purposes. Check our Privacy Policy for details. Click 'Accept all' to let OpenAI and partners use cookies for these purposes. Click 'Reject non-essential' to say no to cookies, except those that are strictly necessary. March 24, 2017 Illustration: Ben Barry We’ve discovered that evolution strategies (ES), an optimization technique that’s been known for decades, rivals the performance of standard reinforcement learning (RL) techniques on modern RL benchmarks (e.g. Atari/MuJoCo), while overcoming many of RL’s inconveniences. In particular, ES is simpler to implement (there is no need for backpropagation⁠ (opens in a new window) ), it is easier to scale in a distributed setting, it does not suffer in settings with sparse rewards, and has fewer hyperparameters⁠ (opens in a new window) . This outcome is surprising because ES resembles simple hill-climbing in a high-dimensional space based onl

We use cookies and similar technologies to deliver, maintain, improve our services and for security purposes. Check our Privacy Policy for details. Click 'Accept all' to let OpenAI and partners use cookies for these purposes. Click 'Reject non-essential' to say no to cookies, except those that are strictly necessary. March 24, 2017 Illustration: Ben Barry We’ve discovered that evolution strategies (ES), an optimization technique that’s been known for decades, rivals the performance of standard reinforcement learning (RL) techniques on modern RL benchmarks (e.g. Atari/MuJoCo), while overcoming