June 2022 Summaries

The text highlights several applications of reinforcement learning (RL) across various domains, including recommender systems, games, production planning, and more. RL is being used to optimize traffic control policies, improve gaming bots' performance, boost wind farm energy production, provide counterfactual explanations for loan decisions, and streamline production scheduling. The Ray Summit conference will feature sessions on these topics, as well as updates on the RLlib library, which is being developed to make RL more accessible and performant for various use cases. Industry leaders are already leveraging RLlib in production across different verticals, and newcomers can attend training sessions to learn more about RL and RLlib.

We've developed an AI- and machine learning-driven IC design orchestration project to tackle the challenges of modern hardware systems, enabling fast and effective design space exploration while reducing computational resources demand. Our approach involves containerizing the digital design flow on a hybrid cloud platform, utilizing Kubernetes clusters for scalable execution and Helm for cloud resource management. We've also developed an ML-driven automatic parameter tuner that leverages Ray and Ray Tune to identify optimal design settings, achieving significant improvements in quality of results compared to manual tuning by expert designers.

Nixtla is an open-source time-series startup that aims to democratize the use of state-of-the-art models in forecasting, leveraging deep learning and efficient statistical models for distributed computing using Ray. Businesses across various industries collect time-series data and seek accurate predictions, with applications ranging from manufacturing and music streaming to supply chain optimization and electricity market prediction. Nixtla's StatsForecast library efficiently addresses the challenge of predicting millions of time series using AutoARIMA, a widely used model that leverages autocorrelations for accurate predictions. By compiling AutoARIMA just-in-time using Numba and distributing computation across multiple cores with Ray, StatsForecast achieves exceptional speed and scalability, enabling data scientists to fit models in under an hour and reducing costs on AWS. The library facilitates tasks such as hyperparameter tuning, model selection, and parallelization, making it easier for users to work with time-series data. With its Ray integration, StatsForecast empowers developers to train more models and scale their forecasting capabilities, providing a powerful tool for businesses seeking efficient time-series forecasting solutions.

Ray 1.13 has been released, featuring improvements in scalable shuffle for large-scale datasets, performance fixes and enhancements to the Ray Serve Deployment Graph API, enhanced stability of autoscaling support through KubeRay, usage stats data collection now on by default, Python 3.10 support, and various other features and enhancements.

The Ray community is coming together at the first-ever in-person Ray Summit, a two-day conference that will cover various topics related to AI and machine learning. The event features keynotes from AI and ML luminaries, as well as breakout sessions, lightning talks, and meetup talks on developer deep dives, use cases across industries, and building scalable machine learning platforms. Sessions include presentations on using Ray for traffic control during the FIFA World Cup, integrating Ray with Spotify's ML Platform, a low-code deep learning platform called Predibase, and leveraging Ray to reduce costs of large model loading and serving. The conference also features talks on infra-less machine learning, large-scale deep learning for production reinforcement learning workloads, and building a serverless resilient graph analysis platform on Ray.