How MuZero, AlphaZero, and AlphaDev are Revolutionizing Computing
The world of artificial intelligence (AI) is constantly evolving, with algorithms becoming more advanced and capable each day. At DeepMind, we are focused on creating AI tools that have a broad understanding of the world and can transfer knowledge between different tasks. Our AI models, such as MuZero and AlphaZero, have already achieved superhuman performance in games and are now expanding their capabilities to optimize data centers and video compression. Additionally, our specialized version of AlphaZero, called AlphaDev, has recently made groundbreaking discoveries in computer science. These advancements are revolutionizing the computing ecosystem and benefiting billions of people and the planet.
Optimizing Data Center Resources
Data centers play a crucial role in managing and processing various tasks, from delivering search results to processing large datasets. At Google, our data center management system, Borg, assigns workloads like a game of multi-dimensional Tetris. However, the manually-coded rules used in Borg are not scalable enough to handle the ever-changing workload distributions. This is where machine learning technologies like AlphaZero come in handy. By training AlphaZero on patterns within data center tasks, it can create optimized and tailored rules to enhance resource utilization.
When we applied AlphaZero to Borg, we observed a significant reduction in underused hardware, up to 19%. This optimization technique has the potential to revolutionize data center resource management and improve overall efficiency.
Next Steps for Video Compression
Video streaming consumes a significant amount of internet traffic and finding ways to optimize video compression can greatly improve user experience. With the help of MuZero, we have successfully reduced the bitrate of videos without compromising visual quality. This enhancement has already improved the YouTube experience for millions of users. Initially, MuZero optimized compression for individual frames, but we have expanded its capabilities to make decisions on grouping and referencing frames during encoding. These efforts have resulted in more bitrate savings and increased efficiency in the video compression process.
Discovering Faster Algorithms
Our specialized version of AlphaZero, AlphaDev, has made a groundbreaking discovery in computer science. It has developed faster sorting and hashing algorithms, which are fundamental processes used in sorting, storing, and retrieving data. These algorithms have wide-ranging impacts, from ranking online search results to improving data retrieval in databases. AlphaDev’s sorting algorithm has shown a 70% increase in efficiency for short sequences and a 1.7% improvement for larger sequences compared to traditional algorithms. The improved hashing algorithm has enhanced efficiency by 30% for data storage and retrieval.
These algorithms have been widely adopted by millions of developers and companies, providing significant time and energy savings across various industries.
The Future of General-Purpose AI Tools
Our AI tools are already saving time and energy for billions of people, but we believe there is much more to come. We envision a future where general-purpose AI tools optimize the entire computing ecosystem that powers our digital world. However, to support these tools, we need faster, more efficient, and sustainable digital infrastructure. Achieving fully generalized AI tools will require additional breakthroughs in theory and technology.
These advancements have the transformative potential to address diverse challenges in technology, science, and medicine. We are optimistic about the future and excited about what lies ahead.
Learn more about sorting algorithms: