Connect with us

Current artificial intelligence programs can defeat humans in such games as chess. However, they still struggle in domains where the rules are unknown or complex, like Atari.

Recently, researchers suggested a new approach that sets a new state-of-the-art result on the Atari benchmark and matches the performance of its predecessor in Go, chess, and shogi.

MuZero Mastering Go chess shogi and Atari without rules

Chess – artistic impression. Image credit: JESHOOTS-com via Pixabay, free licence

Instead of modeling all the environment, the system models only the aspects crucial to the agent’s decision-making process. Specifically, three elements are considered: the value (how good is the current position?), the policy (which action is the best to take?), and the reward (how good was the last action?).

The model can be repeatedly used to improve planning, and increasing the time per move increases playing strength. The novel algorithms make a significant advancement in reinforcement learning. They can be applied in robotics and other real-world environments.

Link to the research article and more information: