Using RLLib in an Enterprise Scale Reinforcement Learning Solution
summary
DeepSim uses RLLib + Tune on a Kubernetes Ray cluster with custom action distributions, loggers and network-config JSONs to run large-scale HPO and export optimized models via Keras/ONNX/TensorRT. An HEV controller trains with PPO for ~6 hours (about 6.5M simulator steps) on Azure HPv2 nodes with four Ray workers and yields a ~3.2% MPG gain over the FastSim baseline. An adaptive cruise controller trains ~10 hours (1M steps) in ANSYS VRXPERIENCE, learning emergency braking in ~5 iterations and robust behavior by ~40 iterations.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 0:56Problem and solution
- 2:15Technical overview
- 5:13Ray platform integration
- 8:06Model deployment
- 10:41Energy management demo
- 18:28Adaptive cruise control demo
related talks
14mConnecting Reinforcement Learning to Simulation Software
17mScaling Reinforcement Learning
20mProject Bonsai: How Microsoft is Using Ray and RLlib Inside a Machine Teaching Service
37mAWS DeepRacer: Train, Evaluate, and Tune
17mReinforcement Learning with Human Feedback in Distributed Systems
30mTransforming Grab’s AI Platform with Ray
27mLessons from Unifying Simulation and RL Training
37mAWS DeepRacer:train, Evaluate, and Tune Your Reinforcement Learning Model