In the early 1990s, Gerald Tesauro (IBM) developed TD-Gammon: a neural net trained with temporal-difference learning by playing millions of games against itself.
The system reached near top-human level and discovered non-obvious strategies. It was an early demonstration of self-play’s power.
Decades later, AlphaGo and AlphaZero inherited the intuition: letting an agent generate experience at scale can beat hand-coded knowledge.
