An opponent that improves with you

Reinforcement learning needs a great many games. Human games are finite, and the supply of recorded Go matches runs out.

So the system played itself.

In self-play, the system generates games by playing against versions of itself. The opponent changes as training changes. It is not guaranteed to sit at the perfect level of difficulty, but it can keep producing new games without waiting for more human records.

The first AlphaGo began with human expert games, then improved through reinforcement learning and self-play. A later system, AlphaGo Zero, changed the architecture and training process. It started with random parameters, learned from self-play using the rules of Go, and did not train on human games. It later beat the version that had defeated Lee Sedol.

The result showed that human games were not required for that later system to reach stronger play. It does not show that human knowledge is always a ceiling, or that removing examples generally improves learning.

Self-play fits a board game because the system can generate many attempts and score each finished game by fixed rules. Most real problems do not provide that setup. You cannot self-play a clinical trial.