Self-Play
A reinforcement learning training paradigm in which an agent generates training data and improves its performance by competing against historical or current versions of itself. The approach is widely applied in the training of board game and video game AI systems.
Timeline (last 90 days)
DeepMind's robot wrestling research conducted in simulated environments demonstrated that agents trained through self-play can spontaneously develop advanced strategies similar to those in human combat, such as feints, counterattacks, and positional control.
In perfect-information zero-sum games, self-play can be proven to converge to a Nash equilibrium strategy.
Sustained self-play is extremely demanding on computational resources, which is a cost challenge for the adversarial co-evolution approach.
All Facts (3)
DeepMind's robot wrestling research conducted in simulated environments demonstrated that agents trained through self-play can spontaneously develop advanced strategies similar to those in human combat, such as feints, counterattacks, and positional control.
50%UnverifiedIn perfect-information zero-sum games, self-play can be proven to converge to a Nash equilibrium strategy.
50%UnverifiedSustained self-play is extremely demanding on computational resources, which is a cost challenge for the adversarial co-evolution approach.
50%