In an article published in the journal Science, Google detailed AlphaZero, an AI-based system that can teach itself to play chess and Go. In each case, she defeated the world champion, demonstrating her ability to learn games from previous matches.
But AlphaZero’s advantage was that it knew the game’s rules in advance and could train before matches began. However, now the researchers have updated the system, now it can learn the game rules during the first game. MuZero’s model predicts the most relevant moves based on the data available at a given moment. She improves her decisions with every move.
The model works in conjunction with the AlphaZero search. Rather than simulating the entire environment with an algorithm, MuZero simulates only those important aspects for making a decision.
Scientists have achieved similar successes because their programs are based on two neural networks at once – computer algorithms that imitate chains of neurons in the human brain. One of these neural networks estimates the player’s current position on the board, and the second uses the first network results. It is she who chooses what to do next and speeds up the calculations.
Over the next several months, DeepMind plans to identify potential commercial applications for MuZero and similar training systems. One of them can be internet traffic. The model can compress video clips and accelerate the performance of the largest video platforms.