Researchers at Elon Musk's startup, OpenAI, think they have discovered the most efficient way to train artificial neural networks: have them compete against each other.
The researchers have created simulations of human sports, like sumo wrestling and soccer, and pitted the AI players against each other. Winning techniques are rewarded through behavioral reinforcement, while the losing AI is encouraged to try different moves.
The 'self-play' method ensures that the AI are always learning, and that their tasks are always the perfect difficulty: facing off against different versions of itself. And, yes, it looks pretty funny.
Topics Artificial Intelligence Elon Musk
