Evolution strategies as a scalable alternative to reinforcement learning
OpenAI Blog
Evolution strategies match the performance of standard reinforcement learning algorithms on benchmark tasks like Atari and MuJoCo environments. The technique achieves this performance while requiring less fine-tuning and being more straightforward to parallelize across distributed systems. This finding suggests evolution strategies could serve as a practical alternative to reinforcement learning for certain optimization problems.
Why it matters
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.
