Deepmind's latest AI doesn't need to know the rulesYou can master the game。 The researchers were published in the company's journal NatureResearch (preprint)A new AI algorithm muzero is reported. Deepmind is famous for beating the go world champion alphago, who has used countless chess games for training.
In 2017, deepmind launched alphazero, which only needs to know the basic rules to self train and master the AI algorithm of the game. Muzero no longer needs to know the rules. It tries to understand the actions allowed by the rules and the rewards brought by specific actions through trial and error. It keeps learning in the process and finds ways to get rewards more easily. By learning rules and improving playing methods at the same time, muzero is more efficient in data utilization than alphazero and other predecessors. Researchers are trying to apply muzero to video compression, autopilot and protein design. This system needs a lot of computing power, but after the training, the decision-making processing power does not need much. It can be operated through smart phones. Many people marvel at children's ability to master knowledge. Linguist Noam Chomsky once advocated that children need to instill grammatical knowledge, otherwise they will be difficult to master language. This view is being questioned more and more.