Competitive self-play
OpenAI Blog
DeepMind trained AI agents through self-play competition to discover athletic skills including tackling, ducking, and catching without explicit instruction. The approach automatically maintains optimal difficulty as agents improve, similar to mechanisms observed in their Dota 2 experiments. This suggests self-play will become a standard component in developing capable AI systems.
Why it matters
We’ve found that self-play allows simulated AIs to discover physical skills like tackling, ducking, faking, kicking, catching, and diving for the ball, without explicitly designing an environment with these skills in mind. Self-play ensures that the environment is always the right difficulty for an AI to improve. Taken alongside our Dota 2 self-play results, we have increasing confidence that self-play will be a core part of powerful AI systems in the future.
