How MuZero, AlphaZero, and AlphaDev are helping optimize the entire computing ecosystem that powers the world of devices
Artificial intelligence (AI) algorithms are getting more sophisticated every day, each designed to solve problems in the best possible way. As part of our effort to build increasingly sophisticated and versatile AI systems, we are working to create AI tools that can make a broader understanding of the world. This makes it possible to transfer useful knowledge across many different kinds of tasks.
Our AI models, AlphaZero and MuZero, based on reinforcement learning, delivered superhuman game-winning performance. Now they are expanding their capabilities to help optimize data centers and video compression. And most recently, a specialized version of AlphaZero called AlphaDev discovered new algorithms that are already accelerating software on the foundations of our digital world.
While these tools are delivering breakthrough efficiencies across the computing ecosystem, early results show the transformative potential of more general-purpose AI tools. Here we explain how these advances are shaping the future of computing and already helping billions of people and the planet.
Optimizing data center resources
Data centers manage everything from delivering search results to processing datasets. Borg manages billions of tasks across his Google, and assigning those workloads is akin to a game of multidimensional Tetris. This system optimizes tasks for internal infrastructure services, user-facing products such as Google Workspace and Search, and also manages batch processing.
Borg manages this workload by scheduling tasks using hand-coded rules. At Google scale, these hand-coded rules are designed as “one-size-fits-all” because they cannot account for the ever-changing variety of workload distributions. best Fits all. This is especially useful for machine learning technologies like AlphaZero. These algorithms can automatically create optimally tuned individual rules that are more efficient for different workload distributions.
During training, AlphaZero learns to recognize patterns in the tasks sent to its data centers, and it can also predict the best ways to manage capacity and make decisions that will yield the best long-term results. I learned.
Applying AlphaZero to Borg, production experiments show that this approach reduces the amount of underutilized hardware by up to 19%, optimizing resource utilization in Google’s data centers. it was done.
The next step in video compression
Video streaming makes up the majority of internet traffic and consumes a lot of data. So finding efficiencies in this process will have a huge impact on the millions of people who watch video every day, both big and small.
Last year, we worked with YouTube to apply MuZero’s problem-solving power to video compression and transmission. MuZero improved the overall YouTube experience by reducing bitrate by 4% without compromising visual quality.
We first applied MuZero to optimize the compression of individual frames in the video. Now, we’ve extended this work to determine how frames are grouped and referenced during encoding, leading to further bitrate savings.
Initial results from these first two steps show great promise for MuZero to become a more generalized tool that helps find optimal solutions throughout the video compression process.
Finding Faster Algorithms
More recently, AlphaDev, a version of AlphaZero, made a new breakthrough in computer science, discovering faster sorting and hashing algorithms. These two fundamental processes are used trillions of times a day to sort, store, and retrieve data.
Sorting algorithms affect how all digital devices process and display information, from ranking online search results and social posts to user recommendations. AlphaDev found an algorithm that improves sorting efficiency for short sequences of elements by 70% and for sequences over 250,000 elements by about 1.7% compared to the algorithm in the C++ library. So when a user submits a search query, AlphaDev’s algorithms can sort the results faster. When used on a large scale, it saves a lot of time and energy.
AlphaDev also found faster algorithms for hashing information, commonly used to store and retrieve data such as customer databases. Hashing algorithms typically use a key (e.g. username “Jane Doe”) to generate a unique hash corresponding to the data value you want to retrieve (e.g. “order number 164335-87”).
Like a librarian using a classification system to find a particular book quickly, with a hashing system the computer already knows what you are looking for and where to find it. When applied to hash functions in the data center 9-16 byte range, AlphaDev’s algorithm improved efficiency by 30%.
Since the release of the LLVM Standard C++ Library’s sorting algorithms (which replace ten-year-old subroutines with RL-generated subroutines) and the abseil library’s hashing algorithms, millions of developers and companies have used these algorithms. Algorithms are used all over the world. Industries such as cloud computing, online shopping, and supply chain management.
Versatile tools to drive your digital future
From playing games to solving the complex engineering problems at the heart of every device, our AI tools save billions of people time and energy. And this is just the beginning.
We envision a future where more general-purpose AI tools help optimize the entire computing ecosystem that underpins the digital world. But supporting these tools will require a faster, more efficient, and more sustainable digital infrastructure.
Many more theoretical and technical breakthroughs are needed to achieve fully generalized AI tools. The transformative potential of general purpose AI tools and how they can be applied to diverse challenges across technology, science and medicine. We are excited about what is to come.
See below for more information on the sorting algorithm.
