Mastering diverse control tasks through world models.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Robotics

Background:

• Current reinforcement learning (RL) algorithms require significant human expertise for new applications.
• General algorithms that learn across diverse tasks remain a fundamental challenge in AI.

Purpose of the Study:

• To present Dreamer, a third-generation general algorithm for artificial intelligence.
• To demonstrate Dreamer's ability to outperform specialized methods across diverse tasks with a single configuration.

Main Methods:

• Dreamer learns a model of the environment and improves behavior by imagining future scenarios.
• Robustness techniques including normalization, balancing, and transformations ensure stable cross-domain learning.

Main Results:

• Dreamer achieves state-of-the-art performance across over 150 diverse tasks with a single configuration.
• Dreamer is the first algorithm to autonomously collect diamonds in Minecraft from scratch, demonstrating farsighted strategy from pixels and sparse rewards.

Conclusions:

• Dreamer offers a general solution for complex control problems, reducing the need for extensive experimentation.
• This advancement significantly broadens the applicability of reinforcement learning across various domains.