Applying Ray and RLlib to Real-life Industrial Use Cases
summary
Claims that RL policies trained with Pathmind + Ray + AnyLogic beat heuristics/optimizers by measurable margins, with one HVAC demo reporting about 75% improvement over a PID baseline. The deployed pipeline exports a policy zip that runs inside AnyLogic; in the demo complaint tickets fall from 217→7 and average energy cost drops from 17 to 11 (results averaged over ~100 Monte Carlo runs). RL training uses standard Ray/RLlib algorithms and is compared directly to heuristics and optimizers.
AI-written summary · about
chapters (16)
- 0:00Introduction and speakers
- 1:22Pathfind’s role and Ray integration
- 2:28Industrial optimization approaches
- 3:50Heuristics vs optimizers
- 6:16Why consider reinforcement learning
- 8:21Industrial RL use cases
- 9:45HVAC demo setup
- 10:29Environment details
- 11:56PID vs RL agent
- 12:38PassMind integration steps
- 14:26PassMind web app walkthrough
- 15:33Configuring reward function
- 17:22Training results overview
- 19:29Policy export and testing
- 20:18Demo results comparison
- 20:38Additional resources
related talks
14mConnecting Reinforcement Learning to Simulation Software
21mAdvanced Ray RLlib: New Features and New Possibilities
20mProject Bonsai: How Microsoft is Using Ray and RLlib Inside a Machine Teaching Service
12mMachine Learning for Adaptive, Self-Optimizing Systems
29mScaling ML Pipelines for Autonomous Driving with Ray
14mRLlib: Scalable RL for TensorFlow, PyTorch, and Beyond
20mInformation Gain Regulation in Reinforcement Learning with the Digital Twins' Level of Realism In RlLib
20mEnergy Optimized Green Light Assist in Varying Traffic Scenarios Using Reinforcement Learning