Scalable AI and reinforcement learning with Ray

Edward Oakes, Peter Schafhalter, and Kristian Hartikainen take a deep dive into Ray, a new distributed execution framework for distributed AI applications developed by machine learning and systems researchers at RISELab, and explore Rays API and system architecture and sharing application examples, including several state-of-the-art distributed training, hyperparameter search, and RL algorithms.

The demands of the modern AI applications continue to grow exponentially, while the improvements in hardware (especially memory capacity) are slowing down, leaving no choice but to scale out these applications. These applications include distributed training, hyperparameter search, and reinforcement learning (RL). These applications have already shown remarkable results, such as disease diagnosis algorithms outperforming medical experts, voice assistants indistinguishable from humans, and AlphaGo beating the world Go champion. However, these applications pose a new set of requirements, the combination of which creates a challenge for existing distributed execution frameworks: computation with millisecond latency at high throughput, adaptive construction of arbitrary task graphs, and execution of heterogeneous kernels over diverse sets of resources. Kristian Hartikainen, Edward Oakes, Peter Schafhalter lead a deep dive into Ray, a new distributed execution framework for distributed AI applications developed by machine learning and systems researchers at UC Berkeley’s RISELab. You’ll walk through Ray’s API and system architecture and explore application examples, including several state-of-the-art distributed training, hyperparameter search, and RL algorithms.