This is where the AI plays against itself to learn and improve. Most modern AIs that are superhuman are all trained on self-play (Poker, Chess, Go etc).

Fictitious Self-Play

In fictitious self-play, the agent trains against an opponent that is also adapting its strategy.

In regular self-play, the agent only learns from its own experiences.

This makes fictitious self-play a more challenging and dynamic learning environment, which can result in agents that are more robust and able to adapt to changing conditions.

Fictitious self-play has been used in various domains, including game theory, robotics, and multi-agent reinforcement learning, to train agents that can learn to cooperate and compete with other agents in complex environments